International Conference on Transportation and Development 2022: Application of Emerging Technologies

Automatic collection of accurate vehicle traffic data, such as vehicle speed, volume, and classification, is critical for the function of smart cities. With the wide availability of traffic cameras, efforts have been made to use AI object detection algorithms for this task. However, the common performance measures used by AI/computer vision communities are mostly concerned with frame level detection performance, which does not directly translate to the accuracy of the traffic data extracted. Therefore, to understand and improve the performance of different AI implementations for traffic data extraction, there is a need for a consistent method that comprehensively evaluates the system performance under diverse conditions. This study seeks to propose a comprehensive evaluation method for AI-based traffic information extraction system. A case study was completed by assessing the performance of a previously developed traffic data extraction system. To demonstrate the benefit of comprehensive evaluation, a headway post-processing method was proposed based on the evaluation results and applied to the system. The case study results showed that the AI-based system produced more accurate results in both non-saturated and saturated traffic flow, tangent roadway geometries, daytime lighting, and lower camera angles, while producing less accurate results in nighttime conditions and at higher camera angles. The headway post-processing method effectively reduced vehicle volume overcounting error to within a 3.8% error regardless of condition. The outcome of this study would support the development of standardized evaluation methodology and public data set to consistently compare the performance of the traffic data extraction system with different AI implementations and drive the advancement in this field.

Safety and security of guided public transportation systems, such as subways, light rails, and, to some extent, APMs, is impacted by various factors. Passengers or people in general may be endangered when falling onto the tracks by driving trains or may constitute a threat to the subway system when willfully entering and damaging the system. The detection system presented in this paper is based on intelligent data sensing and video image analysis (camera, laser combination). It puts a virtual supervisory zone on the platform and the platform track of stations. Several detection modules analyze dedicated parts of the platform and the track and generate alarms in case of dangerous situations. Comprehensive experience has been accumulated for the last 10 years from various test/benchmark installations, e.g., Munich or New York City subway or fully-operational systems, such as the Berlin subway. Functional feasibility and technical stability of the proposed system architecture have been demonstrated; that is, the sensor system can reliably detect any intrusion and just causes a very low number of false detections, which can still be handled without serious impact on train operations.

Lidar based 3D object and tracking is an essential part of autonomous driving. Point cloud data can provide the geometric information among the surrounding objects. However, due to the sparsity and unorderness of the point cloud of pedestrians and cyclists, it is hard to detect and track the pedestrians with a good performance and in real-time at the same time. In order to detect and track the pedestrians and cyclists in real time, we proposed a model, which is based on the YOLOv3 model, a real-time 3D object detector based on the point cloud. First of all, we use the multi-view (MV3D) and Complex-YOLO idea to transform the point cloud into a BEV map. After getting the BEV map, a modified YOLOv3 model is utilized to detect all the pedestrians and cyclists in it; YOLOv3 is an image-based detector, which uses a darknet-53 network and can detect the multi-class objects in a fast speed. Then, with the SORT algorithm, we associate the detections of the pedestrians and cyclists we get in consecutive frames and give them a changeless ID and a 3D detection box. To evaluate our model, we use the widely used KITTI object and tracking objects’ data sets and get an average of 0.86 mAP on the BEV map. And the speed can also reach more than 18 frames per second, which can balance the inference time and disposing efficiency in a good performance.

Transportation accounts for a significant portion of total global energy consumption. Excessive energy consumption usually occurs in urban traffic environments with congestion and travel delays. With the advancement of remote sensing and computer vision technologies, real-time traffic conditions can be monitored. Therefore, sustainable transportation management strategies can be developed to optimize the overall energy and environment performance and reduce congestion and emissions. This work presents a smart traffic monitoring system based on remote camera sensors. Real-time and historical traffic conditions at the US Department of Energy’s Oak Ridge National Laboratory (ORNL) were monitored and analyzed to develop optimal transportation management strategies for sustainability. Computer vision algorithms were developed and applied to process the real-time camera data to obtain complete traffic information across the ORNL campus. Weeks of historical data were collected and processed to analyze the traffic and identify bottlenecks. The proposed traffic monitoring and management approach can be applied and extended to benefit other campuses or urban areas.

The collection of transportation data is crucial to data-driven decision-making. Using artificial intelligence for such data collection efforts could save substantial time and cost and help to improve transportation safety and efficiency. This research study used YOLOv4 object detection and DeepSORT tracking models to count and classify vehicles into 13 FHWA vehicle classes from video footages from existing roadside cameras. YOLOv4 was also used to detect the presence of rumble strips from roadway images and used to create a rumble strip inventory map. The trained vehicle counting/classification model achieves a counting accuracy of ~97% at daytime and ~91% at nighttime. The rumble strip detection model has ~95% accuracy. This study demonstrates that AI model can be trained to accurately count and classify vehicles under various weather, lighting, and traffic conditions, as well as to collect transportation asset data to improve roadway safety.

The main objective of this paper is to develop a low-cost rain detection system using an in-vehicle video camera capable of providing trajectory-level road surface weather information in real-time. The study utilized the Second Strategic Highway Research Program’s (SHRP2) naturalistic driving study (NDS) video data and based on image-based classification techniques. k-means clustering analysis was applied to annotate the extracted images automatically from the video data. To train the detection models local binary pattern (LBP) based images’ features were utilized as classification parameters, and artificial neural network (ANN), random forest (RF), and support vector machine (SVM) were used as classification learners. The overall detection accuracy of the trained RF, ANN, and SVM model was found to be around 99.3%, 98.8%, and 99.8%, respectively. In a connected vehicle (CV) environment, the proposed weather detection system could significantly improve the VSL algorithms, especially on roadways with no weather stations.

This study analyzes the factors affecting motorcycle crash severity in the state of Kentucky while applying machine learning method (i.e., random forest) and deep learning model (i.e., combined principal component-neural network model). Severe motorcycle crashes were the main severity level outcome analyzed in this study and are those crashes resulting in either serious motorcycle injury or fatality. To the authors’ knowledge, these models have been very rarely implemented to analyze motorcycle crashes, especially when it comes to severe motorcycle crashes. Recent five-year motorcycle crash data (2015 to 2019) from Kentucky were used. The random forest classifier was applied to rank each feature’s importance in influencing serious injury or fatal motorcycle crashes. The random forest classifier indicated that collision time, crash location, driver age, helmet use, and vehicle type colliding with the motorcycle were the key features affecting severe motorcycle crashes while yielding 91% prediction accuracy. By testing multiple numbers of principal components, 800 principal components were found to decrease overfitting while still retaining high prediction accuracy. Thus, 800 principal components were used for fitting the neural network model. The neural network showed that driver-related (i.e., age), crash-related (i.e., crash location, collision time, and manner of motorcycle collision), and roadway-related factors (i.e., roadway surface condition) could successfully predict severe motorcycle crashes with 94.2% prediction accuracy. An advanced occlusion-based interpretation of the neural network model also produced a list of features most highly correlated with the model prediction performance. The neural network model result was largely consistent with that of random forest. Nevertheless, deep learning techniques (e.g., the combined principal component-neural network model) could better predict severe motorcycle crashes with higher accuracy compared to machine learning techniques (e.g., random forest). Overall, the study results demonstrated that both machine learning (random forest) and deep learning (combined principal component-neural network model) techniques can be used successfully in identifying those key features contributing to severe motorcycle crashes.

This paper presents an assessment of existing approaches to reconstruct objects via point cloud data with a focus on infrastructure. Light detection and ranging (LiDAR) is a popular remote-sensing method used to collect data for high-resolution models which produces large point cloud data sets. The objective of this research is to facilitate the creation of a digital twin (DT) of a building using point cloud data via automated 3D-model generation and plane classification. This capacity is critical if civil engineering is going to adopt DT at scale for the built environment—much of which is not currently documented in digital form. The 3D reconstruction of large point cloud data sets of infrastructure is an open challenge, in part because of the size of the scan required to create a DT model. An efficient, automated reconstruction method is needed to process LiDAR data. Three reconstruction methods are implemented and qualitatively compared: (1) scan-to-building information model (BIM) approach, (2) oriented point sampling (OPS) algorithm with an emphasis in planar recognition, and (3) RANdom SAmple Consensus (RANSAC) algorithm via the open-source software, CloudCompare. The RANSAC and OPS methods both reduce the size of the point cloud data to improve the computational speed as compared to other traditional methods. A portion of a university campus was chosen as a case study. Ongoing research is focused on creating a DT of a university campus, and this work specifically supports the creation of virtual replicas of large campus infrastructure (i.e., buildings) as one component of a DT. The focus is on the scalability of large data sets from infrastructure projects to compare the accuracy, speed, and computation requirements of each method.

Recent advances in intelligent transportation systems (ITS) and emerging probe-based mobility data sources present new opportunities for effective analysis and improvement of corridor performance. While prior research efforts have looked at the use of specific data sources to address one mode at a time, challenges exist in combining a variety of data sources, each with varying quality and resolution, to understand and improve upon performance of all motorized and non-motorized modes in a unified way. This paper provides an overview of the data types often available to decision-makers and proposes a data model to facilitate the analysis of multimodal data on arterial corridors, including the metadata required to support automated workflows. Two types of metrics are implemented: travel time reliability calculated using probe-based data and transit performance based on automated passenger count data. Our results illustrate the challenges and appropriate interpretations involved in quantifying arterial corridor performance.

Geographic information systems (GIS) and building information modeling (BIM) have many common features and functions for data management and data display and illustration. As part of the Highway Performance Monitoring System (HPMS), the statewide traffic data in Indiana is updated and presented through the GIS method. The attributes of highway location, geometry, and traffic information are expressed and displayed using shapefiles. However, the functionalities of shapefiles in GIS are rather limited in comparison with various BIM features and functions. This paper presents a method to establish a BIM model to replace the existing GIS model in order to improve the effectiveness of traffic information management. The existing GIS traffic information model displays the traffic volumes recorded with portable and temporary traffic data collection devices at the selected sample spots on the Indiana highway network. However, not included in the GIS model are the important traffic attributes recorded with weigh-in-motion (WIM) and automatic traffic recorder (ATR) devices, such as vehicle types, speeds, and axle weights. This paper presents the process of converting the existing GIS model to a BIM-based traffic information management model. The traffic data was expanded to include the WIM and ATR recorded traffic data so that vehicle types, speeds, and axle weights can also be stored and illustrated through the BIM platform. An automatic transformation method was developed and presented in this study. The BIM system has demonstrated strong advantages over the GIS system with respect to traffic data storage, representation, and management.

Occluded person re-identification attracted widespread attention in recent years since it is more consistent with the real-world scenarios compared with other tasks of person re-identification. Researchers proposed a lot of creative works to tackle this task and achieved inspiring progress. However, most of these works severely relied on extra cues from some pre-trained models, such as pose estimator, which made the result of their models sensitive to the estimation error in the extra cues. Moreover, they utilized pose estimator in both training and inference stage, which slowed down the inference speed a lot. In this paper, we propose a novel method named body feature filter (BFF), which is robust to inaccurate pose information from pose estimator by simply transferring score maps of key-points to body part label. Our BFF helps the model filter out noisy information from the occluded regions or the background and focus on the body features of person images. Our method consists of three steps in the training stage. In the first step, we utilize score maps of key-points from pose estimator to get local feature vectors. Then we calculate similarity map in which the value of each location is the similarity score between the local feature vector and the global feature of this location. Finally, we add similarity maps related to the same body part together to generate body part label. In the second step, we design a simple network, BFF, to learn the information from body part label. It generates body part masks to guide the model in knowing where is noisy feature to suppress and where is body part feature to enhance. In the third step, we add a module named body feature refiner (BFR) to further refine the body feature. In the inference stage, we only use body feature filter and body feature refiner without any cues from pose estimator to get final features for person retrieval. The simple architecture of our model makes inference speed faster than that of most other works. Besides, our BFF can be utilized in many person re-identification models, which uses pose estimator. Experiments show the effectiveness of our proposed method.

Pavement markings play a critical role in reducing crashes and improving safety on public roads. Although inspecting pavement markings at the right moment is essential, it is challenging due to the lack of resources for inspections and the geographical distribution of pavement markings. This study proposed a framework that (1) detects pavement markings using deep learning application, (2) effectively stores the information of the detected pavement markings to monitor and manage continuously, and (3) analyzes the condition of the detected pavement markings. The framework adopted the You Only Look Once (YOLO)-v3 algorithm to recognize ten different types of pavement markings. Once the pavement markings are identified, the framework converts GPS coordinates of the recognized pavement markings into Geohash values. Geohash is a multi-dimensional spatial location indexing algorithm that divides regions of the earth. Finally, the framework analyzes the visibility condition of the recognized pavement markings by computing contrast values between pavement markings and surrounding pavements. The framework was tested using selected pavement marking samples in Indiana. The results show that all pavement markings were correctly recognized with precise Geohash IDs and visibility conditions in a reasonable margin of errors. The framework is expected to reduce time and efforts to inspect and monitor various pavement markings.

Weigh-in-motion (WIM) devices classify each passing vehicle into one of the 13 types of vehicles defined by the Federal Highway Administration according to the axle configurations of the vehicle. Accurate vehicle classifications are pivotal for pavement design as they are the main parts of the required design input. However, it is often found that there exist sizable misclassified vehicles in the WIM recorded data. These misclassified vehicles would have major adverse effects on pavement design. This paper presents a study using a machine learning algorithm to identify misclassified vehicles and reclassify them into the appropriate vehicle categories. A video captured near a WIM station was utilized in combination with the WIM recorded data in the machine learning algorithm. The main features of the correctly classified vehicles and the misclassified vehicles were identified and utilized in the training and testing procedures. It proved in this study that the machine learning method could identify the majority of the misclassified vehicles and placed them into the correct vehicle groups. Consequently, the accuracy of the WIM recorded traffic data can be significantly improved, and the quality of pavement design can be greatly enhanced.

Roadside vegetation plays a critical role in the safety and aesthetics of highways and roadways. Well-maintained roadside vegetation aids in soil conservation, improves drainage, stabilizes the slopes, reduces runoff, and improves travel safety. A routine inspection of roadside vegetation is a key requisite for a successful roadside vegetation management system. However, on a regional scale (e.g., county level), the manual routine inspection of roadside vegetation would be extremely labor-intensive and expensive. Therefore, an automated vegetation inspection system is required for the detection and condition assessment of roadside vegetation. To this end, the main objective of this study is to develop an automated inspection system using the state-of-the-art deep learning model to detect roadside vegetation from aerial photography images and classify them based on their quality for maintenance and rehabilitation. The proposed system utilizes the U-Net model to train on a publicly available high-resolution aerial imagery data set from Central Texas and test on images from Texas highways. The results show promising accuracy and precision for roadside vegetation detection and classification. It is expected that the proposed framework will aid transportation agencies in the inspection of roadside vegetation, thereby facilitating proactive rehabilitation decisions.

Pavement performance prediction models are vital to transportation agencies providing decision support for their overall maintenance and rehabilitation (M&R) interventions. Future pavement conditions and rehabilitation priorities have led agencies to the need for developing accurate pavement performance models due to budget allocation issues. The Mississippi Department of Transportation (MDOT) utilizes probability models, which do not include some of the essential parameters such as pavement design, rehabilitation interventions, and traffic. In this study, jointed concrete pavement (JCP) performance models were developed for Mississippi road networks using the artificial neural networks (ANNs) approach to predict the future condition of pavement sections with the effect of M&R actions. A total of 101 JCP sections with 909 data points were used for model development. The developed models successfully characterized the pavement deterioration behavior with a significant agreement between observed and predicted values. Therefore, MDOT can utilize the developed models as a tool for predicting future conditions of JCPs and incorporate the M&R scheduling effectively to prioritize the resources.

Different passenger demand rates in transit stations underscore the importance of adopting operational strategies to provide a demand-responsive service. Aiming at improving passengers’ travel time, the present study introduces a data-driven optimization approach to determine the optimal stop-skip pattern in urban rail transit lines. In detail, first, using the time-series smart card data, we employ a long short-term memory (LSTM) deep learning model to predict the station-level demand rates. Then, we integrate the output of the LSTM model as an input to an optimization model with the objective of minimizing patrons’ total travel time. Moreover, we utilize a heuristic algorithm to solve the problem in a desirable amount of time. Finally, the performance of the proposed models is assessed using real case data. The results suggest that the proposed approach can enhance the performance of the service by improving both passengers’ in-vehicle time as well as passengers’ waiting time.

Trespassing along the railroad rights-of-way is the leading cause of rail-related deaths, with more than 400 annual fatalities in the US. Detection of such events using video camera systems is critical for railroad safety improvements while challenging due to the immense labor costs required for the processing and monitoring of a large number of streamed video files. The objective of this study is to develop a deep learning based automatic trespassing detection process using streamed railway surveillance video data, which will assist in identifying, characterizing, and alerting potential trespassers in the region of interest (ROI). In this research, the You Only Look Once (YOLO) v4 and Deep SORT are implemented for trespassing detection and tracking. The result shows that all trespassers in the video were detected without false negative or false positive. The detection results can be applied to the trespasser detection and alerting systems for sending a timely alert to the trespasser, sending the trespassing information to railroad safety officials, and saving trespassing event video data to the database. It is anticipated that this work could assist railroad agencies to improve railroad security based on real-time detection results.

Roadway safety remains a critical issue due to the disproportionate number of crashes in rural areas. According to the National Highway Traffic Safety Administration (NHTSA), from 2007 to 2016, the fatality rate in rural roads has been more than two times higher than in urban areas. This discrepancy in fatalities reveals the urgency to improve the roadway safety conditions in the rural areas to achieve transportation equity. While most rural roadway safety related studies have focused on major facilities, few efforts have been made to investigate the crash injuries on the low-volume roads. This study investigated the potential influential factors, such as roadway geometric characteristics, environmental conditions, human behaviors, traffic conditions, etc., on the injury severity of crashes occurred on low-volume rural roads. Four models, including ordered probit (OP), multinomial logit (MNL), artificial neural network (ANN), and random forest (RF), were trained, tested, and validated using five years of Washington State crash records on low-volume rural roads from 2013 to 2017. It was found that the two machine learning models (ANN and RF) performed better than the two statistical models (OP and MNL), and the RF model had the best performance in predicting crash injury severities. The results also showed that variables such as grade percentage, degrees of curvature, shoulder width, drivers’ gender, roadway width, head on crash, pedestrian/cyclist involved, young driver, truck involved, etc. have significant impact on the crash injury severities on low-volume rural roads.

Caltrans’ performance measurement system (PeMS) collects traffic data from thousands of detectors. These detectors could malfunction resulting in less than 100% to 0% observation at certain vehicle detection stations (VDS). Thus, this study aims to predict traffic flow when no data is available on faulty detectors using machine learning techniques such as linear regression and random forest. The entire length of I-680 South was considered using flow data over two weeks. The results of the prediction of flow show that using the eigenvectors from principal component analysis in a linear regression model gives very low accuracy. Meanwhile, simply using the select features from the data results in better performance which is then improved when the output from k-means clustering was added as a feature. Meanwhile, using random forest gives the best performance compared to the previous models but with a greater model training time.

Recognizing travel patterns of the urban traffic is crucial for improving the performance of urban traffic control, thus mitigating the negative impacts of traffic congestions on our daily life. In recent years, tons of license plate sensors are deployed on urban traffic networks to monitor motorized traffic by recording the license plate number of passing vehicles at intersections. The large-scale license plate sensing data (LPS) do contain valuable vehicle origin-destination (O-D) information, which can be utilized for unveiling travel patterns of urban traffic. This study developed the structure of travel pattern recognition, which consists of three components: data pre-processing, feature extraction, and pattern recognition modeling. The architecture was tested and validated using the LPS data sets from Nanjing, China. The results showed six clusters has low sum of square errors (SSE) and high Calinski-Harabasz (CH) index, which can explain the changes of travel behaviors the urban traffic especially in the morning O-D trips.

Traffic density has increased exponentially in recent years, and real-time traffic flow monitoring has become necessary for areas with high-density traffic flow. It is essential to measure the number of vehicles in the road networks in order to optimize the traffic management. However, estimating traffic volume efficiently and accurately is still a challenge. This paper applies a one-dimensional (1D) and a three-dimensional (3D) glass fiber-reinforced polymer packaged fiber Bragg grating (GFRP-FBG) sensor, which is durable with relatively low cost to collect strain signals. These strain signals are produced once vehicles pass through the GFRP-FBG sensors, and the number of vehicles is able to be obtained by counting the strain wavelength collected through Luna’s optical sensing interrogator. With the real-time traffic on the I-94 freeway, this study validated the effectiveness of the 3D GFRP-FBG with the video that is recorded simultaneously.

Based on the sensing task in intelligent vehicle infrastructure cooperative systems (IVICS), this research mainly studied the end-to-end multi-target tracking method of multi-sensor fusion. Aiming at the fusion of point cloud data generated by lidar and camera vision, this paper studied the data level fusion method, 2D-3D fusion 3D object detection method, and 3D multi-object tracking method, respectively. Aiming at the vehicle object detection in intelligent transportation environment, this paper proposed a new extraction and fusion algorithm based on four levels of point cloud eigenvalue, and verified the algorithm model based on KITTI data set. In this study, by improving the PointFusion object detection algorithm, 3D multi-object detection was provided for traffic scenes, and the heterogeneous Pointnet ++ network was introduced to improve the performance of point cloud feature extraction. In this paper, the multi-level features of point cloud and the global features of image were extracted, respectively. Finally, the fusion of feature level provided a 3D multi-object detection model. The densefusion architecture was transformed into an end-to-end algorithm by data level fusion, and the performance of the model was optimized by bin based parameter classification estimation, multi-level feature extraction of point cloud, and focal loss of class balance.

Traffic situation awareness in urban environments is an essential technology for applications such as collision detection and congestion prediction. Correlational research has received much attention due to the significance of traffic security and intelligence. On this issue, the combination of multiple sensors (e.g., lidar, radar, monocular, and binocular cameras) promotes the perception and decision-making process of vehicles, based on an integrated information about the environment. The paper proposed an end-to-end algorithm utilizing multi-sensor data fusion to forecast collision risk of vehicles. The algorithm was implemented in two stages. In the first stage, both radar and camera data was taken as input. Then, motion features of vehicles were extracted through object detection and tracking algorithms. Afterwards, we performed trajectory prediction and motion state estimation using data obtained from last step. During the process, collision risk for each vehicle was estimated according to indicators like time-to-collision (TTC) and minimum-future-spacing (MFS). In the second stage, the extended Kalman filter (EKF) was applied to smooth original predictions, thus improving the degree of confidence in the data. In the algorithm implementation, algorithms of object detection, tracking, and trajectory prediction were connected with necessary modifications, reaching a balance between accuracy and efficiency. The algorithm was tested with nuScenes data set to prove its effectiveness in actual traffic scenarios. Experiments showed that our algorithm could output acceptable collision risk of vehicles over time, thus providing information for further inference about traffic situations.

The safety of pedestrian crossing has always been an important issue in the field of traffic safety. However, at present, vehicle road coordination early warning still needs to be improved. Therefore, a millimeter-wave radar vehicle-road collaborative pedestrian warning device using edge computing modules is proposed, which can detect pedestrians in real time, efficiently and accurately, then alarms drivers. The device includes five modules: radar data acquisition module, edge computing module, data transfer module, cloud server, and data display module. The system architecture is different from the traditional architecture. It combines artificial intelligence technology and the Internet of Things, and delegates a large amount of computing work to the edge computing module to obtain better system performance. Moreover, a RSU (road side unit) is designed to store, process, and upload the data processed by the edge computing module near the intersection in real time, instead of directly transmitting the data to the cloud server. In the actual deployment, this device is proved that it is able to have a good and stable detection in dark night environment and complex weather.