CHAPTER 2

Review of the State of the Art and State of the Practice of Machine Learning in Transportation

State of the Art in ML Applications for Transportation

Within the last two decades, ML has gained significant momentum due to the availability of large data, faster computing, and more importantly, the proliferation of deep learning (DL) methods capable of discovering highly complex patterns. Deep learning methods (e.g., Convolutional Neural Networks (CNNs), recurrent neural networks (RNNs), and more recently transformers) have been applied in various fields including computer vision, natural language processing (NLP), bioinformatics, and robotics, and, in some cases, exceeded human expert performance (e.g., DeepMind’s AlphaGo). DL methods can learn and improve their performance with access to larger volumes of data. While ML/AI methods have a promising future, the path to success is rarely a straight line. There are risks and challenges to be overcome. Explainability, interpretability, reliability, social impacts, bias, and fairness of ML methods are being debated and potential solutions are explored. Replacing existing ‘legacy’ systems that are time-tested and proven reliable with ML/AI-based alternatives will not be straightforward as decision-makers may not be aware of their benefits and risks. In some cases, ML/AI solutions just may not be mature enough for deployment by transportation agencies. Despite the risks and challenges, ML/AI has been making significant inroads into all major sectors of the economy over the last decade.

In the transportation sector, a prime example is automated vehicles (AVs) which are powered by advanced DL methods that process large sensor data in real-time to detect and classify static and dynamic objects in the driving environment. The private sector is investing billions of dollars to improve AV technologies to enable full automation and driverless robo-taxis and transit shuttles. On the public sector side, state and local Departments of Transportation (DOTs) are starting to explore ML methods so that they can capitalize on the solutions offered by AI/ML to improve the safety, mobility, efficiency, and sustainability of multimodal transportation systems.

This section presents a review of the transportation literature on ML methods and the types of application areas these methods have been applied. First, a brief overview of ML is presented which is followed by a synthesis of the ML literature in transportation. Later in the chapter, the state of

the practice of ML is presented and specific examples of how transportation agencies are exploring ML-based solutions are provided.

A Brief Overview of ML

ML is broadly defined as the capability of a machine or computer to learn (from data and experience) without explicitly being programmed. Learning in the ML context involves discovering patterns and relationships (e.g., factors leading to higher crash risks), making useful predictions (e.g., bus arrival time), exhibiting intelligent behavior (e.g., generalizing from an example), and mimicking human characteristics and skills (e.g., creative art, conversation, driving a car). For machines to accomplish such complex tasks, numerous types of algorithms and methods have been developed some of which are inspired by how a human brain works (e.g., artificial neural networks), and others are founded on statistical learning theory (e.g., support vector machines (SVMs)). In either case, all ML algorithms require data – images, time series data, text, survey data, etc. – without which learning is not possible. ML algorithms fundamentally learn from data through a process called model training. Hence, more and better-quality data lead to better ML models. On the other hand, if the data are biased, the resulting model will also be biased.

Most ML methods are developed to solve specific problems and tasks such as classification, regression, feature extraction or dimensionality reduction, system optimization and control, etc. Given a specific task and its supporting data, the ML process involves selecting an appropriate model and training the model based on the data to the extent possible so that it meets certain performance criteria (e.g., the error rate is acceptable). Typically, there are numerous candidate ML models that may be considered for the task. Therefore, a trial-and-error process is commonly employed to determine which ML model will be the best choice. The end result is an ML algorithm for future computations.

All ML methods have parameters, the values of which must be learned from the available data through the model training process. The number of these parameters can vary radically depending on the complexity of the model. For example, the k-nearest neighbor (k-NN) algorithm has only one parameter – the number of training examples that are closest to the input data point for which a prediction is to be made – whereas advanced DL models can include billions of parameters. For model training, special computational (iterative) methods are available for determining these model parameters (e.g., stochastic gradient descent for deep neural networks). Model training is typically the most computationally intensive step in the ML pipeline.

Depending on how learning from data is accomplished and what type of data are provided as the input, ML methods are categorized into the following types:

Supervised Learning (SL): In supervised learning, the ML model is supplied with both the input data and the correct output (i.e., labels) that the ML algorithm is expected to produce from the input data. Given a set of input-output pairs, the ML model is trained such that it learns how to map inputs to outputs. The training involves selecting a model, varying its parameters, and evaluating its performance. The model that produces the most accurate mapping based on the training data is typically deemed the best. The model accuracy can be measured through various loss functions (e.g., mean square error for regression problems and cross-entropy loss for classification problems.) Supervised learning is one of the most common types of ML. However, labeled data may not be available or may be cost-prohibitive which limits the applicability of supervised learning methods.

Unsupervised Learning (UL): In this case, the data are not labeled (e.g., the correct answer is unknown) and the ML model attempts to find underlying patterns in the input data that may not be known a priori. For example, the model may group the input data into distinct clusters, find useful relationships or associations between input variables, and reduce the high number of dimensions to fewer for better data visualization, analysis, and compression.

Reinforcement Learning (RL): In RL, the learning takes place through interactions (e.g., trial and error) of an agent or agents with an environment. The agent takes actions and subsequently observes their

consequences (through a predefined reward function) and collects data from the environment. The environment can be a computer game (or any game), traffic simulation, or real-world as long as the actions taken by the agent have an effect on the system and/or the reward function. The agent learns the best set of actions to take or a policy to maximize its reward. RL-based methods are commonly used in control theory, simulation-based optimization, and multi-agent systems, among others.

In addition to the three basic ML paradigms listed above, there are other nuanced types of learning that define how the problem is being solved and typically pertain to DL methods. Some of these additional learning types that may be applicable to transportation problems are briefly summarized.

Semi-supervised learning refers to the cases where only a small portion of the data are labeled but the ML model attempts to make an effective use of all the data by employing both supervised and unsupervised modeling techniques.

Self-supervised learning refers to applying SL methods to data that are not labeled but instead, labels are generated from the input data itself. For example, given a large set of images, a certain portion of the image may be removed from the input and considered as the label or output. The same idea can be applied to a text where a section of the sentence is omitted and considered as the label. The ML model is then trained to predict the omitted portion of the data. Autoencoders and generative adversarial networks (GANs) are examples of such self-supervised ML methods.

Transfer learning refers to models that are previously trained and subsequently applied to a related task. For example, a DL model trained on a large dataset may be fine-tuned on a smaller dataset for solving a classification problem.

Continual learning (lifelong learning or online learning) refers to the types of models where the learning is continuous as new data become available. The model adapts itself and incrementally learns new information from the streaming data that may potentially drift and exhibit new behavior. Continual learning is an active fundamental research area in the ML/AI field.

ML in Transportation Research

Breakthroughs in ML and DL are creating tremendous opportunities to automate many challenging tasks and to improve transportation system efficiency and safety. For example, connected and automated vehicles (CAVs) and numerous sensors within the multimodal transportation systems are increasingly generating a massive amount of data that cannot be processed fast enough with traditional statistical approaches to extract actionable insights. There is an extensive body of academic literature focused on the applications of ML to a broad range of transportation problems. Within the last two decades, thousands of articles have been published in academic journals and presented at transportation conferences.

To synthesize the large body of academic literature on ML applications in transportation, a two-pronged approach is taken in this project. First, recently published survey papers are reviewed to identify key trends and document specific transportation problems being addressed. Second, articles published in the Transportation Research Record (TRR) are analyzed to quantify trends in the use of ML methods over the last 20+ years. A summary of these TRR articles is presented in the next subsection. The section below provides highlights from a few selected review papers which is then followed by a summary table listing common application areas and problems for which ML methods have been explored.

Haghighat et al. (2020) provides a review of DL applications in Intelligent Transportation Systems (ITS) based on papers published between January 2015 and October 2019. They cite 265 papers in their review and categorize the ITS-related papers into six different areas: Traffic characteristics prediction, traffic incidents inference, vehicle identification, traffic signal timing, ride-sharing, and public transportation, and visual recognition tasks (e.g., traffic sign and light recognition). For these applications, where available, the authors report the level of accuracies attained by DL methods in comparison to various baseline methods as reported in the reviewed papers. The authors conclude that DL “undeniably achieved better results as compared to existing techniques” especially for traffic state prediction, vehicle identification, and visual

recognition tasks. They argue that significant improvements need to be made in the use of DL models for other areas such as incident detection, signal timing, ride-sharing, and other public transportation concerns. They also highlight the importance (or lack) of benchmark datasets that are needed for facilitating a fair comparison of competing models (Haghighat et al. 2020).

Abduljabbar et al. (2019) provides an overview of AI applications in transportation. They argue that AI can be greatly beneficial for ITS applications since in such problems the system dynamics and user’s behavior are too difficult to model and predict. They explore the role of AI in different applications of planning, public transportation, intelligent urban mobility, and CAVs. For each of these categories, they study different algorithms implemented and compare their relative strengths and challenges. Among all applications, they report that AI methods have been successful in alleviating the effects of uncertainties and gaps within the data which is a great challenge for traditional methods. A detailed study of the limitations of AI/ML techniques is also presented in their study. For example, they report that forecasting under unexpected events and adverse weather conditions is not currently done accurately enough for practical purposes and urge researchers to develop more reliable weather and incident-responsive algorithms. The perception of neural networks as black boxes, computational complexity, and the possibility of bias introduction during manual human labeling are among other limitations listed. They believe that DL will be the future of AI and provide an estimate of the economic, social, and business value added with this projected AI usage.

Chung et al. (2021) presents a review of the applications of AI/ML in logistics and transport. They first introduce a summary of review papers on emerging technologies in logistics from 2015 to 2021. Their study also classifies each paper by application categories of pattern recognition, prediction, and classification. For each application, they provide a detailed report of the technical difficulties and challenges encountered during the development of ML techniques. Moreover, their paper explores the role of using autonomous systems in different transportation optimization problems and reports significant benefits in many applications such as in public transportation and last-mile deliveries. They conclude that applications of smart technologies and ML will be critical in the near future and list a few possible significant directions they believe will gain momentum such as using blockchain technologies in logistic digitalization.

Boukerche et al. (2020) survey and compare ML and statistics-based methods in short-term traffic flow prediction. They consider the ARIMA (autoregressive integrated moving average) model and its derivatives, and exponential smoothing as representatives of statistical models and provide a detailed comparison of them with ML algorithms such as k-NN, SVM, LSTM (long short-term memory), and RNN. They also review other algorithms such as Kalman filtering, and hidden Markov machine as an independent category. The advantages and disadvantages of each of these algorithms for short-term flow prediction are studied and discussed. Their review emphasizes that ML approaches have been used more frequently in recent years thanks to their better model flexibility, higher model adaptability, and stronger non-linear feature mining ability. However, they mention model complexity and data limitation as the strongest challenges these ML models currently face.

Clearly, ML methods have been applied to a variety of transportation applications. Table 1 shows a broad list of application areas, and for each one of these ten areas several examples or problems for which ML methods have been developed and proposed in the academic literature. These examples are not meant to be exhaustive but rather representative of the types of problems that are commonly addressed in the literature by ML. A few review or survey papers for each application area are included as sample references. It should be noted that many more references can be listed, and researchers are continuing to publish new reviews as the ML methods evolve.

Table 1 Example types of problems being solved by ML methods for ten application areas based on the literature review.

ML-related Papers Published in the Transportation Research Record

This section presents an analysis of ML papers published in the Transportation Research Record (TRR). This analysis is based on the citation data for all TRR papers published over 23 years, from 1999 to 2022. The data include their titles, abstracts, keywords, authors, and other bibliographic information. After downloading the bibliographic data, performing a quality check, and eliminating papers without abstracts (e.g., forewords in the TRR books) a total of 19,675 papers are found. As manually reviewing these many papers is time-prohibitive, a text mining approach is selected to identify ML-related papers among the 19,675 TRR papers. For this purpose, a set of key phrases shown in Table 2 are used. Initially, a larger set of key phrases were identified but they were refined and reduced to the list shown in the table as additional key phrases did not result in identifying more ML-related papers. Furthermore, it should be noted that this list is not meant to be exhaustive to capture all ML methods. However, it is created to capture the great majority of ML methods, especially those that have been dominant in recent years (e.g., deep learning). For example, logistic regression or linear regression is not included (The word “regression” appears at least once in 1,358 papers.) The goal in creating this list is twofold: (i) to study the trend in ML-related publications over the years; and (ii) to find out which ML methods are more commonly being applied in the transportation research field.

Table 2 List of key phrases used in identifying ML-related papers.

A text mining program was created in the statistical programming language R to count the number of times the key phrases in Table 2 appear in the abstract, title, and keywords reported by the authors in each TRR paper. The presence of at least one of the key phrases corresponding to a given ML method in a paper is assumed to be an indicator that the subject paper makes use of the given ML method. This resulted in

identifying 1,115 papers as ML-related out of the initial corpus of 19,675. While this approach may be simplistic, the results looked generally accurate based on randomly checking some of the subsets of papers assigned to an ML category. For example, all papers with the “reinforcement learning” keyword were found to be indeed about RL-based methods. It should be noted that this approach does not result in one-to-one mapping of the papers to the 10 ML methods listed in Table 2. Each ML-related paper is mapped to at least one ML method, but some are mapped to more than one category. Nevertheless, most of the papers, 817 out of 1,115, were mapped to only one ML method. This is expected as some papers involve applying and developing different types of ML methods.

Figure 3 shows the total number of all papers as well as those identified as ML-related across the years. The secondary axis shows the percentage of ML-related papers out of the total count. It is evident that starting in 2017 there is a remarkable increase in the number of ML-related papers published in TRR. The percentage of ML papers jumps from an average of 4% in the years before 2017 to about 15% in 2021. To put this transition in perspective, it is interesting to note that in the early 2010s DL became popular after DL-based methods achieved remarkable accuracies in international data competitions (e.g., ImageNet) and DL tools became widely available for training large models (e.g., the initial release of TensorFlow in 2015). A rapid increase in ML-related papers after 2017 is observed in other fields as well (Pugliese et al. 2021).

Figure 4 shows a breakdown of ML papers in TRR by the ML methods listed in the first column of Table 2. The “ML/AI” category is not included as the key phrases for this category are generic and can include all types of ML methods. The trend line indicated by “NN-DL” includes those papers in which NN key phrases occur, but “DL” key phrases do not in their abstracts, titles, and keywords. From the figure, starting in 2017, a rapid increase in the number of DL papers is evident. Except for GA (and to some extent for the “Fuzzy” category), papers in all other categories have been increasing significantly within the last few years. It should be noted that NLP, RL, and DL have become popular recently and there are close to zero papers in TRR on these methods before 2017. The surge in DL papers starting in 2016 observed in Figure 4 is consistent with studies on ML trends and publications in other fields (Haghighat et al. 2020).

Figure 5 shows the distribution of the identified ML papers in TRR by application area and travel mode. The 1,115 papers were manually categorized into the ten application areas listed on the vertical axis based on the information from the paper title and abstract. For simplicity, each paper is assigned to a single application area even though this may not be the case for all papers. For example, papers on vehicle classification are considered in the operations category even though vehicle classification information is used in pavement design as well. Also, papers on highway-railroad crossing issues were considered to be in the highway mode. A few papers were excluded as they were about transportation research (e.g., studies on trends in transportation research (Boyer et al. 2017; Das et al. 2017)) and were not related to one of the listed application areas. In addition, a few papers were explicitly about multimodal freight operations and therefore they were not assigned to one of the modes and were left out of the data for Figure 5. Nevertheless, these exceptions were not large enough to affect the overall distribution of papers by application area. As expected, the great majority of papers (about 92%) are from the highway mode whereas the rail mode is a distant second represented in about 5% of the papers. In terms of application areas, the top five are operations, planning, infrastructure/asset management, safety, and transit. The common types of problems being addressed by ML methods in these application areas are listed in Table 2.

Advancements in DL Methods

Within the last decade, there has been remarkable progress in deep learning (DL) that enabled the ML field to achieve unprecedented advancements in AI applications, transforming industries and research by enhancing capabilities in image and speech recognition, natural language processing, autonomous systems, and predictive analytics. This progress has been largely driven by advances in algorithms, computational power, and the availability of large datasets. The decade saw significant improvements in neural network architectures, with Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) leading to breakthroughs in image and speech recognition, respectively. CNNs excel in processing data with a grid-like topology (e.g., images), while RNNs are designed to handle sequential data (e.g., text). The emergence of DL frameworks like TensorFlow, PyTorch, and Keras (see Chapter 5 of this report for more information on ML tools) has democratized access to powerful DL tools. These frameworks offer high-level interfaces, making it easier for researchers and developers to design, train, and deploy complex neural network models. Also, advances in hardware, particularly GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units), have considerably accelerated the training of DL models.

The introduction of innovative neural network architectures also represents significant milestones in the evolution of DL methods. For example, first introduced in 2017 by a team at Google in the research paper “Attention is All You Need” transformers revolutionized how machines understand human language and interpret images (Vaswani et al. 2017), marking a significant leap in natural language processing (NLP) and computer vision fields. Unlike previous models that processed data sequentially, transformers use self-attention mechanisms to weigh the importance of different words in a sentence (or different parts of the input data), enabling parallel processing and reducing training times. This mechanism enables the model to focus on specific (relevant) parts of the input data. This is critical for understanding the context in sentences or images, as it allows the model to remember and emphasize relevant information without the computational costs or limitations of older methods. Transformers architecture is the foundation of many popular ML models like Google’s BERT (Bidirectional Encoder Representations from Transformers) and OpenAI’s GPT (Generative Pre-trained Transformer) series.

For image interpretation, transformers treat the image as a sequence of pixels or patches, analogous to how they treat a sentence as a sequence of words. This approach enables the model to understand and interpret complex patterns, relationships, and contexts within images. The introduction of architectures

like Vision Transformer (ViT) segments an image into fixed-size patches, linearly embeds each of them, adds positional embeddings, and feeds the sequence of embeddings into a standard transformer encoder. By doing so, it can effectively learn spatial hierarchies and feature representations that are critical for tasks such as image classification, object detection, and semantic segmentation. Moreover, transformers have facilitated the development of multimodal models that can process and relate information across different types of data, such as text and images, within a single framework. This has opened new avenues for applications like image captioning, visual question answering, and cross-modal information retrieval, where understanding the interplay between visual elements and natural language is essential.

One of the important features of transformers is their ability to be trained on vast amounts of unlabeled data. This is done through a process called semi-supervised learning, where the model learns from both labeled and unlabeled data, significantly reducing the need for manually labeled datasets. Once trained, these models can be fine-tuned for specific tasks using a smaller set of labeled data, a method known as transfer learning. BERT is a prime example of this approach, offering advanced capabilities for a wide range of language tasks, from question answering to language inference. New models are being developed to extend the capabilities of the original transformers architecture. Some applications of the transformer-based DL models in transportation are discussed below.

Recently, transformer-based models have been implemented by researchers in the transportation field. Aside from using transformers for NLP applications (Tsai et al. 2022), these models have also been applied to a broad range of ITS applications. For example, Dong et. al (2021) proposed an explainable autonomous driving system using an image transformer model. They offer an end-to-end autonomous driving system using a self-attention mechanism to map visual features from camera images to guide potential driving actions with a corresponding explanation. Xu et al. (2022) developed a model to predict vehicle trajectories in congested urban traffic. They show that their model significantly outperforms LSTM in both accuracy and training time. Yan et al. (2021) proposed a transformer model to learn spatiotemporal features of traffic and have shown significant accuracy improvement over both traditional models such as ARIMA and deep learning models such as DCRNN and FC-LSTM at a fraction of the training cost. From traffic flow estimation and prediction (Yan, Ma, and Pu 2021; C. Chen et al. 2022; Fang et al. 2022; J. Zhang et al. 2021; Li and Lasenby 2022), car-following trajectory prediction (M. Zhu et al. 2022), lane changing detection (Jun et al. 2021), traffic road object detection (W. Liang et al. 2022; Manzari et al. 2022; Yu et al. 2020; Jiaming Zhang et al. 2022), accident detection and prediction (W. Liu et al. 2021; Yifan et al. 2022; H. Liu et al. 2021), distracted driver detection and assistance (Hong Vin et al. 2021b; J. Liang et al. 2022; Huiqin et al. 2021; Hong Vin et al. 2021a), to pavement condition classification (Y. Chen et al. 2022; Xiang et al. 2022), transformer models are increasingly used in various applications, and a significant improvement over the previous ML methods is reported in these publications.

Another type of DL architecture that has found a wide range of applications is the generative adversarial networks (GANs). Originally proposed by Ian Goodfellow and his colleagues in 2014, GANs consist of two neural networks, a generator and a discriminator, that are trained simultaneously through adversarial processes (Goodfellow et al. 2014). GANs are used to generate artificial samples that are indistinguishable from real data as much as possible. GANs are widely used for generating realistic images, enhancing photographs, creating art, and more. The potential of GANs in various transportation applications has been explored by many researchers (Lin et al. 2023). Specific applications of GANs in transportation include autonomous driving, where they are used to generate synthetic/perceptual data for different driving conditions (Santana and Hotz 2016); traffic flow forecasting where GANs are used to address incomplete or sparse traffic data (N. Wang et al. 2023); predicting the deterioration of bridges where they are used to generate synthetic images of bridges showing various stages of wear (Bianchi and Hebdon 2021); pavement crack detection where GANs are used to generated images of different crack types (Han et al. 2024); and floodwater detection on roadways where GANs are used to create realistic videos of floods (Lamczyk et al. 2022). These, and many other examples, show that GAN architectures have the potential to generate realistic large synthetic data consistent with the actual conditions. These

large diverse datasets help improve the generalization ability of models used in transportation applications.

Machine Learning State of the Practice at Transportation Agencies

The following sections summarize the state of the practice of machine learning (ML) in the context of state departments of transportation (DOT) and other transportation agencies based on available online references (e.g., reports, articles, slide decks, case studies, press releases, etc.). However, it can be difficult to find information regarding transportation agency involvement with ML in published documents since many agencies may be informally experimenting with ML (either directly or with partners) and, therefore, not ready to publish results. Additionally, if the agency is not using federal support to develop its ML capabilities, then it may not have reporting requirements.

Even if references are available, it can be difficult to assess the level of agency involvement in ML development. In some cases, agencies may be partnering with nearby university researchers who are leading ML development with varying levels of agency involvement in the process. For example, researchers often reach out to agencies for permission to use agency data (e.g., closed-circuit television (CCTV) feeds) for ML research and development.

In other cases, agencies may be deploying ML capabilities with vendors but not be privy to the ML specifics. Conversely, some agencies may think their vendors are using AI/ML based on their marketing but, in reality, they may be using the terms generously. In either situation, it can be difficult for agencies to know the details of the ML due to the proprietary nature of vendor capabilities.

Given these potential limitations and considerations, this chapter attempts to summarize available insights on the state of the practice of ML with transportation agencies. The examples mentioned are not meant to be exhaustive, but rather indicative of the general state of the practice. Overall, there seem to be quite a few agencies considering, leveraging, and implementing ML. The following sections summarize ML trends at agencies and the state of the practice of ML by broad transportation areas.

ML Trends at Agencies

This section summarizes high-level trends gleaned during the literature review process with respect to the state of the practice at agencies of common types of ML, transportation areas, procurement/collaboration, ML development, data, supporting technology, and funding and risk.

Common Types of ML: Supervised ML with a human-in-the-loop seems to be the most prevalent form of ML being developed or implemented at agencies. Some popular areas include computer vision (usually with deep learning) and short-term prediction tasks. For example, a number of agencies are using computer vision for object, incident, and/or pedestrian detection. Additionally, edge computing is becoming increasingly common. For example, more than half of respondents to an AI for ITS Sources Sought Notice indicated they made use of edge computing in their AI-enabled applications (Vasudevan et al. 2022).

Common Transportation Areas for ML: TSMO (Transportation Systems Management and Operations) appears to be the most popular area for ML at agencies, but it is also the broadest transportation area. There also appears to be a lot of research emphasis on safety-related applications, such as hazard detection and driver behavior monitoring. Asset management is an area that is seeing increased interest and utilization as well.

Procurement/Collaboration: Of the agencies developing and/or implementing ML, most seem to procure ML capabilities from vendors and/or leverage their university partners.

ML Development: Few ML developers, whether they be within the agency or from a consultant, vendor, or university partner, are developing new ML solutions from scratch. Instead, developers commonly use available pre-trained baseline models (e.g., YOLO for object detection) and other open-source tools as a

starting point. For example, most TSMO agencies procure transportation management system (TMS) applications from vendors, and those vendors will likely use components from open-source tools provided by major platforms and other third-party AI software providers to integrate AI into existing applications and in developing new AI-supported applications (Gettman 2019).

Data: Agencies are using a variety of data sources to support their ML applications. For example, many appear to be using their existing agency data sources when possible (e.g., camera, weather, detector, etc.) as well as acquiring or purchasing additional data sources (e.g., INRIX, Waze, etc.).

Supporting Technology: Data storage, computational, and software resources are becoming increasingly important foundational elements to support ML applications. These elements are commonly sold as a subscription service. Procuring software on a subscription basis is an emerging practice for most DOTs (Gettman 2019). Pricing models vary based on computing cycles, data storage size, the frequency of analytics, and other metrics, making it difficult to estimate how much an application will cost (Gettman 2019).

Funding and Risk: Some agencies are using federal grant funding to kickstart their ML programs. For example, the U.S. DOT’s Advanced Transportation and Congestion Management Deployment (ATCMTD) Program has helped support a number of ML-focused deployments. Overall, “larger organizations (from a TSMO perspective, this would imply larger State departments of transportation (DOT) and the largest cities) generally have more tolerance for downside investment risk with the potential for substantial at scale benefits if successful” (Gettman 2019).

State of the Practice of ML by Transportation Area

This section focuses on examples of ML research, testing, and deployment in the context of state DOTs and other transportation agencies broken down by broad transportation area. The following transportation areas are discussed based on the availability of relevant literature: vehicle automation, safety applications and driver behavior, planning applications, TSMO applications, asset management, commercial vehicle and freight operations, transit operations, and traveler information and accessibility. TSMO applications are further broken down into those for work zone management, traffic incident management (TIM), road weather management, traffic estimation and prediction for decision support, vulnerable road user detection, intelligent traffic signals, and other miscellaneous operations. This state of the practice review was conducted in summer-fall 2022.

Vehicle Automation

Automated Vehicles (AVs) have long been a horizon of the future of transportation for transportation agencies and the private sector. The most cited reference for automated vehicles is the Society of Automotive Engineering (SAE) J3016 standard, which defines six levels of driving automation: (0) no driving automation, (1) driver assistance, (2) partial driving automation, (3) conditional driving automation, (4) high driving automation, and (5) full driving automation (SAE ORAD Committee and ISO TC204/WG14 2021). AVs function through the orchestration of many different systems of environmental sensors, algorithms that decide the behavior of the vehicle, and the vehicle systems that move the vehicle. ML is very frequently used as the technology backbone for sensing and motion planning. It consolidates sensor data from sources like cameras, LiDAR, and radar systems. Machine vision, a subfield of ML, infers what is happening around the vehicle, and detects the state of other roadway actors (e.g., pedestrians, vehicles, bicyclists) as well as traffic signals, signs, and debris. It even attempts to predict what road actors are likely intending to do in the next several seconds. By synthesizing this information, ML systems then conduct motion planning by selecting from an output of a set of driving actions (Fowler 2021).

Although much of the cutting-edge work in this area is being done by technology companies and auto manufacturers, agencies have begun testing and engaging with automated vehicle deployments as well. For example, NHTSA launched the AV-Test Tracking Tool, which collects and presents in a user interface

information about automated vehicle tests that states and companies have validated on public roads or in laboratory settings (National Highway Traffic Safety Administration 2020). NHTSA has also been involved with hands-on testing of automated driving. The Naturalistic Study of Level 2 Driving Automation Functions (Russell et al. 2018) conducted a study of the interactions between real-world drivers and commercially available vehicles with some automated driving capabilities.

FHWA’s cooperative driving automation (CDA) CARMA program is an open-source platform meant to accelerate the research and development of CDA concepts. The platform takes advantage of ML techniques such as machine vision and data fusion. It enables researchers and engineers to develop and test CDA concepts having to do with infrastructure, automated vehicles, and other road users (FHWA 2022c). The program intends to build the foundation for the adoption of CDA that works across transportation infrastructure and vehicle make. It examines the impact of automated driving not just on typical traffic but also on freight operations.

Several state transportation agencies and the Department of Motor Vehicles have set up programs that issue permits to manufacturers of autonomous vehicles to test their vehicles on public roads. Typically, the manufacturer must demonstrate that these tests will pass a set of safety protocols before they are issued these permits. Examples of states with these types of programs include California (California DMV, n.d.), Massachusetts (MassDOT, n.d.), Vermont (Vermont Agency of Transportation 2020), and more.

Some agencies are exploring pilots centered around automated electric shuttles. The Colorado Smart Cities Alliance, a statewide partnership of government, business, and non-profit organizations established by the Denver South Economic Development Partnership, launched Autonomous Vehicles Colorado (AvCo) in 2021. The project is a deployment of driverless electric shuttles in three locations. The shuttles helped to connect communities and navigate traffic using advanced sensors (Svitak 2022). The Florida Department of Transportation (FDOT), Pinellas Suncoast Transit Authority (PSTA), and a company called Beep launched an AV pilot program in St. Petersburg. The vehicle drives along a pre-determined route and uses LiDAR sensors and GPS tracking to maintain the path during operation (Pinellas Suncoast Transit Authority (PSTA) 2020). The city of Greensville County, South Carolina is using a $4 million grant to deploy a system of driverless taxis to improve access for disadvantaged and mobility-impaired residents (Singer 2017).

Some agencies are taking the innovation challenge approach to the deployment of AV solutions within their state. For instance, the Minnesota Department of Transportation (MnDOT) has issued the CAV Challenge request for proposals. They receive ideas from interested parties and review them to ensure that the ideas align with state goals, benefit the public, and fill research gaps. In later phases, after panel approvals, they allocate funding to some of the contestants. The program has earned the “3^rd most innovative procurement” from the National Cronin Awards for contract excellence (MnDOT 2022).

Safety Applications and Driver Behavior

Safety is a key priority for state and local transportation agencies, and some are researching the use of ML to support various roadway safety goals. For example, the University of Utah, in research performed for the Utah DOT, developed ML-based safety models (i.e., artificial neural network and support vector machine) that model crash frequency and severity along the I-80 corridor to help evaluate road safety improvements for variable speed limit (VSL) deployments (Azin and Yang 2021). Additionally, in a project supported by the FHWA Exploratory Advanced Research (EAR) Program and designed to assist state DOTs, researchers from the Road Ecology Center at the University of California, Davis used AI processes to analyze and detect animals in camera trap images and worked with state DOT and natural resources agency staff to teach them how to use the developed tools (FHWA 2022a). In another project supported by the FHWA EAR Program, the Palo Alto Research Center (PARC) developed ML tools to process data from several datasets—including the Second Strategic Highway Research Program (SHRP2) naturalistic driving study (NDS) dataset, data on the physical characteristics of frequently traveled roadway sections, weather

data and video logs, and video, radar, and image data from Chicago intersections—to detect safety issues that might have been missed otherwise (FHWA 2022a).

The City of Bellevue in Washington conducted a safety-focused project beginning in August 2019 titled “Video-based Network-wide Conflict and Speed Analysis to Support Vision Zero in Bellevue (WA)” (City of Bellevue, WA 2021). The city partnered with vendors to support their data collection and safety analysis capabilities and installed high-definition traffic cameras at 40 strategic intersections (ITS Deployment Evaluation 2021). With data collected from these high-quality video feeds and proprietary AI software for advanced video analytics, they identified traffic volumes, road user speeds, and near-crash traffic conflict indicators. The results allowed practitioners to identify the most problematic intersections and better understand the nature of contributing factors to crashes [(City of Bellevue, WA 2021) & (ITS Deployment Evaluation 2021)].

3D point cloud and high-resolution video data generated by unmanned aerial systems (UASs) are becoming increasingly popular sources of data for agencies to support safety initiatives (Gettman 2019). ML tools, such as neural networks and other image processing algorithms, can be used to extract important features from these UAS-generated data (Balali et al.2017). For example, the North Carolina DOT and North Carolina Highway Patrol have researched the use of UASs to perform crash reconstruction data collection (Gettman 2019). In 2017, they conducted a study by simulating a head-on crash on a divided highway in a controlled environment and compared the data collection and processing times using the traditional manual method used by the highway patrol collision reconstruction unit to the new method using three UASs operated by pilots. While the highway patrol unit took nearly two hours to collect the data, the UASs took only 25 minutes, which would have resulted in an estimated $9,300 in savings had the simulated crash occurred on I-95 (Drake 2018).

Driver behavior plays a major role in roadway safety since an estimated 94% of serious crashes can be attributed to human error, according to the National Highway Traffic Safety Administration (NHTSA) (S. Singh 2015). The FHWA EAR Program is supporting a variety of research projects to analyze driver behavior using ML-based video analytics (FHWA 2022a). For example, researchers at the University of Michigan developed automated video-processing algorithms to detect and classify different road user behaviors and actions (e.g., talking on a cell phone, searching for an address). Additionally, researchers at Carnegie Mellon University developed an automated real-time system to analyze drivers’ emotional states as well as their level of distraction and fatigue. In a third and final example, researchers at the University of Wisconsin-Madison developed an open-source software platform that could quantify driver distraction and engagement.

Since ML use cases may overlap across multiple transportation areas (e.g., safety, planning, freight, etc.), the remaining sections may include additional safety-related ML examples.

Planning Applications

Given ML’s ability to learn patterns in large historical datasets, it may lend itself well to planning applications. For example, in research supported by the Maryland DOT, researchers from the University of Maryland developed a variety of ML models that can be used to assist with decision-making in project planning and construction (Y. Zhang, Cutts, and Xu 2021). For example, the researchers developed and tested reinforcement learning models for drilling project schedule estimation using historical data records. Some of the ML models have been integrated into existing work processes at the Maryland DOT State Highway Administration (Y. Zhang et al. 2021). Additionally, researchers from the Missouri University of Science and Technology used predictive deep learning methods to support flood evacuation planning and routing for the Missouri DOT (Corns et al. 2020). Their research developed forecasting tools capable of assessing the water level rate of change in high-risk flood areas to help determine evacuation routing and detours to mitigate the potential loss of life during flash floods. Transportation safety and/or disaster planners could use these results to support evacuation planning and routing.

The remaining sections (e.g., Asset Management) discuss the state of the practice of other planning-related ML applications.

TSMO Applications

According to the Federal Highway Administration (FHWA) Office of Operations (FHWA 2022b), Transportation Systems Management and Operations (TSMO) is a set of strategies that focus on operational improvements that can maintain and even restore the performance of the existing transportation system before extra capacity is needed, with the overall goal of maximizing the performance of existing transportation facilities. According to a report by the U.S. DOT Intelligent Transportation Systems (ITS) Joint Program Office (JPO) (Vasudevan, Townsend, Dang, et al. 2020), AI could be applied at the system, technical, and/or operational levels for TSMO. For example, applications could range from specific TSMO programs such as work zone management, traffic incident management, or road weather management while operational tactics could include service optimization, such as variable speed limits or adaptive traffic signals.

Currently, TSMO may be the most active area for ML at agencies. However, since it is such a broad category, this section has been broken down further into a handful of the most common ML application areas that agencies seem to be pursuing within TSMO.

Work Zone Management

Some agencies are exploring the use of ML to support their work zone management programs. For example, the Iowa DOT in partnership with the Institute of Transportation (InTrans) at Iowa State University is using ML to support real-time performance monitoring of work zones. Specifically, InTrans used ML to develop an application that identifies slow and stopped conditions and sends text alerts along with camera images to operators during sustained traffic delays so they can have immediate awareness of the situation and respond quickly (Iowa Department of Transportation 2021). Iowa DOT’s focused attention on their work zone management program leverages the partnership with InTrans for specialized ML expertise and has generally accelerated the adoption of emerging technologies (Iowa Department of Transportation 2021).

Additionally, according to their TSMO Strategic Plan Update (AECOM 2021b), the Texas DOT is planning to use Cooperative Automated Transportation (CAT) data to support a variety of AI and ML applications, including those for smart work zone management. Specifically, Texas DOT’s 2021 report prepared by AECOM on CAT integration for TSMO (AECOM 2021a) mentions that CAT technologies, such as connected and automated vehicles, will contribute additional sensing and computing power to the existing infrastructure-based ITS devices, which could support new ML use cases for smart work zone management. For example, ML can be applied to quickly detect the onset of congestion, a developing queue, or an incident requiring rapid response.

Finally, a research team from the Volpe National Transportation Systems Center developed a video-processing tool that uses ML to train neural networks to identify and classify roadway features and driving conditions in the SHRP2 NDS dataset (Rittmuller 2022), which was a research project supported by the FHWA EAR Program (FHWA 2022a). In phase I of the project, the team focused on detecting and mapping work-zone features, such as barrels, cones, and signs. In phase II, they trained neural networks to detect traffic signals, signal states, weather events, and roadway weather conditions. These extracted features can be added to the NDS database to enable safety researchers and other stakeholders to use the processed data for their ML use cases.

These are examples of how ML is being used, planned, and researched by (or in support of) agencies for work zone management based on the review of available sources.

Traffic Incident Management

Several DOTs are using or exploring the potential of ML to support traffic incident management. Many DOTs, including but not limited to Nevada (Gettman 2019), Florida (Gettman 2019), North Carolina (Kolar 2021), Utah (Rekor Systems 2021b), Louisiana (Rekor Systems 2021a), and Missouri (Rekor Systems 2022), are using vendors for ML-based incident detection capabilities. For example, both the Nevada and Florida DOTs are using a proprietary, cloud-hosted, ML-based software subscription service as a supplement for incident detection. The software fuses data from a variety of sources including radar detectors, loop detectors, and closed-circuit television (CCTV) cameras to train a neural network to recognize scenes as “incidents,” “not incidents,” or “incidents may be likely to occur.” Since using the service, both DOTs have reported improvements in incident detection times and reductions in crashes from positioning highway patrol assets strategically and providing advanced warnings to motorists via dynamic message signs (Gettman 2019). Southern Nevada has seen an average reduction in incident response times of 12 minutes since using the software (Southern Nevada TMC 2019), and both the Nevada and Florida DOTs have seen an approximate reduction in crashes, particularly secondary crashes, by 17 percent (Gettman 2019). Additionally, according to vendor press releases [(Kolar 2021); (Rekor Systems 2021b); (Rekor Systems 2021a); (Rekor Systems 2022)], the North Carolina DOT, Utah DOT with the Utah Department of Public Safety, Louisiana Department of Transportation and Development, and Missouri DOT are overseeing pilot deployment programs with a vendor’s AI-based technology to aid in incident detection and crash prediction using a variety of data sources, including data from connected vehicles and mobile apps.

Missouri DOT’s integration of the vendor’s AI-based technology for crash prediction and incident detection is part of their larger ATCMTD deployment of predictive analytics for the traffic management platform, which integrates real-time and historical data sources and frequently uses AI/ML, to support decision-making in the near-term with the overall goal of improving roadway safety (Allmeroth and Intaratip 2022). Missouri DOT piloted the capability in a heavy construction area on I-270 in the St. Louis District and evaluated the algorithms for accuracy every quarter beginning in August 2021 (Allmeroth and Intaratip 2022). Missouri DOT is working with one of its vendors to improve the accuracy of the crash prediction results to 15% - 20% of the total crashes predicted (Allmeroth and Intaratip 2022). To increase accuracy for both the incident detection and crash prediction algorithms, Missouri DOT is integrating additional connected vehicle data feeds, installing 74 static cameras to collect traffic data, and pursuing St. Louis County 911 Computer Aided Dispatch (CAD) data (Allmeroth and Intaratip 2022). Based on updated data from the same source from October 2023, the predictive algorithm appears to have improved some over time, predicting 80% of true incidents in February 2022, 95% of true incidents in April 2023, and 86% of true incidents in July 2023.

Other DOTs are partnering with research or academic institutions to support their traffic incident management efforts with ML. For example, the Georgia Tech Research Corporation conducted a feasibility study (Guin et al. 2020) for the Georgia DOT of a video-based automatic incident detection (AID) technology on an approximately 16-mile stretch of I-475 during a 3-month period in 2018. The study proposed a clustering ML framework to identify alarms for high-impact incidents requiring immediate attention from emergency and TMC responders. Additional filtering layers were added on top of the ML algorithm to help eliminate the majority of false, unverifiable, and noncritical alarms since false/noncritical alarms (i.e., false positives) are a common challenge for ML-based incident detection algorithms. Additionally, the Iowa DOT partnered with Iowa State University to develop an incident detection system called Traffic Incident Management Enabled by Large Data Innovations (TIMELI) that uses neural networks and supporting software to classify images from traffic surveillance cameras across the state in near real-time (Gettman 2019). Overall, Iowa’s TIMELI system is focused on improving incident detection time, particularly in rural areas where camera surveillance may be available, but it may take a while for highway patrol to receive notifications. The Tennessee DOT has also partnered with an academic institution, Vanderbilt University, to improve its traffic incident management capabilities with ML (Baroud

et al. 2021). The researchers developed a framework that uses synthetic resampling, clustering, and data mining techniques to efficiently forecast the spatial-temporal dynamics of accident occurrence (i.e., predict the risk of highway incidents), even under sparse conditions, using various datasets related to roadway geometry, weather, historical accidents, and traffic. Simulation results showed up to a 19 percent average improvement in response time.

Finally, at least one DOT is working with a mix of private sector and university partners to develop ML capabilities for incident management. The Washington State DOT, as part of their ATCMTD deployment, is developing a Virtual Coordination Center (VCC) to enable real-time data sharing and coordinated response across multiple agencies to clear roadway incidents quickly (FHWA 2020b). The VCC is a dashboard platform that combines data in near real-time from agencies across Seattle to help operators identify incidents requiring collaboration from multiple agencies (e.g., large crashes, fires, fatalities, etc.) (Vasudevan et al. 2022). While the first wave of their project focused on setting up the integrated data infrastructure and developing shared situational awareness, the next wave of the project will include developing a supervised ML algorithm to automatically flag/identify potential cross-agency incidents after sufficient labeled data has been created in the new platform (Vasudevan et al. 2022).

Road Weather Management

Some agencies are testing or researching ML to support road weather management. For example, the Wyoming DOT has been considering ML to support its various road weather management efforts. First, the University of Wyoming conducted a study for the Wyoming DOT (Ahmed et al. 2021) using the SHRP2 NDS video data that included using ML to identify road weather conditions (e.g., clear, rainy, snowy, and foggy). Second, the Wyoming DOT used the Pikalert® System, which was developed by the National Center for Atmospheric Research (NCAR) with funding and support from the USDOT (National Center for Atmospheric Research, n.d.), as part of their Connected Vehicle Pilot (ITS Joint Program Office, n.d.). The system processes weather and vehicle data to identify precipitation, pavement, visibility, and blowover conditions and issues actionable weather alerts (e.g., high wind blowover hazards) and forecasts to TMC operators (Young, Welch, and Siems-Anderson 2019). The system is available as open source for agencies to use (“OSADP/Pikalert-Vehicle-Data-Translator-” 2020).

Additionally, in work sponsored by the FHWA and performed by a consulting firm, the integrated modeling for road condition prediction (IMRCP) system was demonstrated at the deployment site in the Kansas City area with the Kansas City Scout and Missouri DOT (Sanchez, Neuner, and Gonzalez 2020). The IMRCP system draws input from traffic, weather, and hydrological data sources (with atmospheric and hydrological conditions drawn from National Oceanic and Atmospheric Administration/National Weather Service sources). State and local agencies provide additional specialized data, such as pavement temperatures. Using these data, the system estimates current road weather (e.g., icy, wet) and traffic conditions and forecasts future conditions, of which some are ML-based predictions. The IMRCP provides prediction data on web-based maps and enables users to select layers for road conditions, weather, and alert data [(Garrett et al. 2017) & (Pisano 2018)]. According to the Missouri DOT website (Allmeroth and Intaratip 2022), the IMRCP platform was fully operational in February 2022, but it is not fully used during the non-winter season since the technology is mainly focused on supporting winter operations.

Traffic Estimation and Prediction for Decision Support

Traffic estimation and prediction for decision support, short-term in particular, is a large area for ML research, testing, and implementation. Some DOTs are sponsoring or engaging in research in this area. For example, the Louisiana Department of Transportation and Development sponsored work conducted by the Louisiana Transportation Research Center to explore non-traditional methods of obtaining vehicle volumes (Codjoe, Thapa, and Yeboah 2020). Although ML was not the focus of this effort, the researchers assessed StreetLight InSight® and Streetlytics, produced by StreetLight Data and Bentley Systems respectively, regarding their ability to provide traffic volumes, including annual average daily traffic (AADT) volumes.

Some of these data (e.g., StreetLight AADT 2018 V3) were generated using ML trained with real-world data, Navigation-GPS, and Location-Based Services (LBS) data. Additionally, the Florida DOT-sponsored work conducted by Florida International University (Hadi et al. 2019) to assess the applicability, feasibility, and effectiveness of data, analysis, modeling, machine learning, and simulation approaches to support decision-making for off-line and real-time operations of integrated corridor management (ICM), including exploring the use of ML to categorize traffic patterns and to automate the process of updating the signal timing plans during non-recurrent conditions. Finally, a research team from Carnegie Mellon University is developing AI-based algorithms using historical multi-source traffic data to predict non-recurrent traffic conditions in large networks at least 30 minutes ahead and to proactively recommend operational management strategies in real-time, which is a research project supported by the FHWA EAR Program (FHWA 2022a). The research team intends to conduct case studies in one small municipality network and two large-scale regional networks.

Some agencies are planning to develop and use ML for decision support. For example, the Texas DOT plans to develop AI and ML tools to assist in making more informed decisions regarding traffic, safety, incident, emergency, and asset management, according to their 2021 Statewide TSMO Strategic Plan Update (AECOM 2021b).

Finally, a few agencies are planning, testing, or implementing ML-based applications, such as traffic prediction, for ICM decision support as part of their ATCMTD deployments, including Virginia DOT (FHWA 2020a), Tennessee DOT (Singer 2020a), and Delaware DOT (Delaware DOT, n.d.). For example, Virginia DOT is in the planning and procurement phase for their AI-based decision support system that will use ML to predict when incidents and congestion are likely to occur and when an event requires a coordinated multi-agency response, with the overall goal of enhancing traffic incident management in support of ICM in Northern Virginia and Metropolitan Fredericksburg (Virginia DOT 2022). Additionally, Tennessee DOT is designing and building its AI-powered decision support system for ICM that will use a variety of management strategies on the I-24 Smart Corridor, including variable speed limits (VSL) and lane control systems (LCS) (Work 2022). Finally, Delaware DOT is developing its own AI-based integrated transportation management system (AI-ITMS), which includes a variety of ML components. In a presentation as of January 2022, they have implemented 2 ML use cases (i.e., short-term traffic flow prediction and traffic incident detection and classification), have developed and are testing 3 ML use cases (i.e., incident response decision support, vehicle reidentification, and machine vision), and are planning one use case for the future (i.e., system self-learning with reinforcement learning) (Donaldson et al. 2022).

Vulnerable Road User Detection

A number of agencies are engaging with their university partners to research, develop, and demonstrate computer vision applications for vulnerable road user (VRU) detection, including New York City DOT (ITS Deployment Evaluation 2022), the City of Pittsburgh (Kocamaz et al. 2016), and Caltrans (Pourhomayoun 2020). For example, researchers at New York University’s Connected Cities for Smart Mobility toward Accessible and Reliable Transportation (C2SMART) University Transportation Center (UTC) leveraged existing New York City DOT CCTV traffic cameras and computer vision for pedestrian detection with plans to expand their approach to other detection use cases, such as detecting parking occupancy, monitoring bus lane usage, identifying illegal/double parking, and assessing pedestrian demand at transit stops (ITS Deployment Evaluation 2022). Additionally, researchers from Carnegie Mellon University’s Robotics Institute developed a computer vision-based cyclist and pedestrian detection, tracking, and counting method with the goal of providing actionable insights for government officials and advocates, as part of a project in partnership with the City of Pittsburgh (Kocamaz et al. 2016). Similarly, in work sponsored by Caltrans with the goal of improving pedestrian and bicyclist safety, researchers from the Mineta Transportation Institute at San José State University developed a computer vision-based cyclist and pedestrian detection, tracking, and counting system and evaluated the system on videos captured by actual traffic cameras in the city of Los Angeles (Pourhomayoun 2020).

Some agencies are procuring computer vision capabilities from vendors for VRU detection. For example, the City of Denver, as part of their ATCMTD deployment’s “Safer Pedestrian Crossings for Connected Citizens” goal area, contracted with a Boulder-based visual intelligence company to automate pedestrian detection and notification to improve pedestrian safety (Boulder AI 2020).

Intelligent Traffic Signals

Intelligent traffic signal systems are another technology area that agencies are exploring for TSMO. For example, the City of Pittsburgh, as part of its ATCMTD deployment, is expanding its network of adaptive signal controllers along eight priority corridors they are calling “Smart Spines” to provide connectivity between different neighborhoods and the downtown central business district (Maloch 2021). The project seeks to leverage existing and emerging technologies to create a multimodal advanced signal system able to detect different roadway users and prioritize their movements based on context and real-time traffic patterns (Maloch 2021). Scalable Urban Traffic Control (Surtrac) is one example of a next-generation intelligent traffic signal control system that uses AI and is being deployed in some locations, including in Pittsburgh, PA; Quincy, MA (a suburb of Boston); and Portland, ME (Vasudevan, Townsend, Dang, et al. 2020). Another intelligent signal application that could use ML/AI capabilities is emergency vehicle preemption (EVP). For example, the Sacramento, CA suburb of Rancho Cordova piloted an AI-enabled EVP system from LYT on 27 signalized intersections on two major streets, according to a company case study (Lee 2022a). Results from a before-after study of the pilot indicated that the technology decreased vehicle travel time on average by 6.6% and reduced the average time to reach an incident by 34 seconds of the six pilot vehicles (Lee 2022a).

Other Misc. Operations

There are a variety of other TSMO application areas that agencies may be considering using or are using ML. One of these areas is with respect to vehicle occupancy detection. For example, the Regional Transportation Commission of Southern Nevada, as part of their ATCMTD deployment, is planning to deploy high-occupancy vehicle detection using emerging technologies and data analytics to support ICM strategies along a critical US-95 corridor in the Las Vegas metropolitan area (Hoeft 2019).

Only a handful of the major TSMO application areas for ML are mentioned in this section based on the availability of literature/references connected to agencies. Other TSMO application areas that are not included here but could have active ML examples from agencies include dynamic tolling, etc.

Asset Management

Asset management includes a broad spectrum of activities and strategies for effectively managing infrastructure assets to maximize their value and service life while minimizing associated costs and risks. From condition assessment (e.g., pavement monitoring) to inventory management (e.g., surveying road signs), ML could provide significant value to state DOTs by automating various processes and supporting decision-making. According to Yinhai Wang, chair of TRB’s Standing Committee on Artificial Intelligence and Advanced Computing Applications, transportation asset management will be an important application of AI in transportation. “I believe there is a great potential to use AI for infrastructure risk and deterioration assessments [like predicting which bridge is getting close to the end of its useful lifespan]” (National Academy of Sciences 2022). Many University Transportation Centers (UTCs) are conducting research into this topic, and some agencies have launched pilots that make use of ML for asset management. A few examples are mentioned here. For many other great examples of how ML is being used for asset management, please see the section in this report on “Surveys with State DOTs.”

Unmanned aerial systems (UAS) provide an opportunity for inspection of pavements, bridges, and other transportation infrastructure using video feeds, photography, and LiDAR. 3D point cloud data and high-resolution video generated by UAS are ripe for analysis with AI. As of March 2018, 35 of 50 U.S. State DOTs have a UAS program, with 20 of the 35 using UASs for daily operations and 15 in the research phase

(Gettman 2019). However, it is unclear how many of those agencies are making use of AI to analyze the data collected with UASs.

The city of Dublin, Ohio wanted to improve its pavement maintenance process and conduct a road assessment on its 278 centerline-mile road network. Previous assessments required months to do data analysis because man man-hours were spent converting the data to work with their pavement management system. They partnered with AI company RoadBotics to collect data and conduct the assessment. Data collection was completed quickly with a smartphone camera and the platform used AI to assign conditional 1-5 ratings for every 10-foot section of the roadway, with results displayed on an interactive map (RoadBotics 2022).

In May 2021, the Utah Department of Transportation (UDOT) announced that it will conduct a pilot with an AI company called Blyncsy. It will use a solution called Payver to assess more than 350 road miles in Salt Lake County. This application uses dash-cams to gather data and then uses proprietary machine vision algorithms to deliver insights. It houses this information on a platform that supports data collection, aggregation, and analysis. UDOT hopes to use it to frequently access necessary road condition data and prioritize and manage its resources (Stone 2021).

Commercial Vehicle and Freight Operations

AI and AVs hold significant potential for increased efficiency of commercial vehicle and freight operations. As part of FHWA’s ATCMTD grant program, the Virginia Port Authority is developing a proof of concept for using autonomous trucks to access shipping terminals and deliver and receive containers. The initiative seeks to prepare the Port of Virginia to service the next generation of high-tech autonomous or semi-autonomous trucks. The goal is not to integrate this type of technology in the short-term but to begin understanding the future requirements from an operational and infrastructure standpoint to deploy these types of trucks (Roberts 2021).

Another deployment is DriveOhio’s I-70 Truck Automation Corridor project. DriveOhio is a part of the Ohio Department of Transportation that serves as its hub for smart mobility. The project intends to demonstrate three levels of technology: SAE Level 2 automated trucks, SAE Level 4 automated trucks, and truck platoons, which although not an AI technology itself, can synergize well with AVs. The project will begin with a roadway audit of I-70 to evaluate the automation readiness of the freeway, give recommendations on changes to infrastructure owner-operators, and develop open-source software for evaluation of the road’s AV readiness with the intention of sharing this code with other agencies in the country. They also plan on authoring an AV Readiness Guidebook that presents findings and lessons learned from the roadway audit (DriveOhio 2020).

Transit Operations

A few agencies have procured or are planning to procure AI/ML capabilities from vendors to support transit operations. For example, the city of East Palo Alto deployed an AI-enabled Transit Signal Priority (TSP) system from LYT on four major signalized along a busy corridor, according to a company case study (Lee 2022c). Results showed a reduction in northbound intersection delays by 45% and southbound intersection delays by 19%, which translates to 18% and 7% reductions in travel time for northbound and southbound, respectively (Lee 2022c). Additionally, the City of San José recently decided to expand their deployment of AI-enabled TSP from LYT from the piloted 20 traffic signals to 142 signals to prioritize the Santa Clara Valley Transportation Authority, according to a company press release (Lee 2022b). According to LYT’s CEO, their cloud-based TSP uses ML to predict the optimal time to grant the green light to transit vehicles at the right time (Lee 2022b). The Heart of Iowa Regional Transit Agency, as part of their Complete Trip – ITS4US Deployment project, plans to use Uber Transit, which likely develops its prediction models using historical performance data and ML techniques, for improved Estimated Time of Arrival (ETA) prediction accuracy (Mishra and Ramsey 2022).

Traveler Information and Accessibility

A few agencies have begun to explore the possibility of ML to improve traveler information and/or accessibility and are planning to use it in their deployments. For example, the Metropolitan Washington Council of Governments, as part of their ATCMTD deployment, will leverage real-time big data, AI, and advanced computing to provide personalized, timely, and accurate travel information to all residents and visitors in the service area and provide incentives to minimize congestion, energy use, and emissions [(Singer 2020b) & (Sheehan 2021)]. Additionally, the Pinellas County Department of Public Works, as part of their ATCMTD-supported “Pinellas Connected Community” project, will integrate data from third-party providers and video analytics into their smart city data platform, will deploy predictive analytics for incident and congestion risk predictions, and will use connected vehicle technology to alert motorists of traffic jams and help them avoid accidents along SR 60 and US 19 [(Singer 2020c) & (Arasteh, n.d.)].

To support accessibility, the Atlanta Regional Commission Complete Trip – ITS4US Deployment project called “Safe Trips in a Connected Transportation Network” is leveraging advanced transportation technology solutions, including ML and predictive analytics, to support safe and complete trips, with a focus on accessibility for those with disabilities, aging adults, and those with limited English proficiency (Wakhisi et al. 2022). Specifically, the Space-Time Memory platform will be used to process traffic volume and speed data from multiple monitoring and modeling sources, track and predict multimodal network performance, and predict evolving route conditions using traditional and ML techniques (Wakhisi et al. 2022).

Some agencies are researching the potential of ML to improve the accuracy of travel time information. For example, in research supported by the Illinois DOT, the Illinois Center for Transportation at the University of Illinois at Urbana-Champaign developed an enhanced travel time prediction model for northeast Illinois expressways using ML and multiple data sources, including loop detectors, probe vehicles, weather conditions, geometry, roadway incidents, roadwork, special events, and sun glare (Mohammadian, Taghipour, and Parsa 2020).

Chatbots

Chatbot-related applications are becoming increasingly popular, especially after the release of ChatGPT in November 2022 (after the time of this state of the practice review). Some agencies are using chatbots for natural language questions and answering to support 511 systems and traffic engineering operations (Gettman 2019). For example, the Metropolitan Transportation Commission (MTC) Bay Area 511 now has an Alexa skill to pass through requests for 511 information for the same phraseology that works with their interactive voice response module (Gettman 2019). Additionally, the City of Surprise in Arizona, which is a suburb of the Phoenix metropolitan area, developed a Google Assistant interface in 2018 for its adaptive traffic control system (Gettman 2019). The chatbot system allows traffic engineers to query status data using voice commands through Google Assistant installed on a phone or other device (Gettman 2019). While these are just a few operations-related examples of chatbots used at agencies, plenty of new examples are likely to have emerged since this state-of-the-art practice review in summer-fall 2022.