A deep learning approach for real-time crash prediction using vehicle-by-vehicle data

Analyzing the heterogenous effects of factors on high-range speeding likelihood of taxi speeders: Does explainable deep learning provides more insights than random parameter approach?

The random parameters Generalized Linear Model (GLM) is frequently used to model speeding characteristics and capture the heterogenous effects of factors. However, this statistical approach is seldom employed for prediction and generalization due to the challenge of transferring its predefined errors. Recently, the emergence of explainable AI techniques has illuminated a new path for analyzing factors associated with risky driving behaviors. Despite this, there remains a gap that comparing results from machine and deep learning (ML/DL) approaches with those from random parameters GLM. This study aims to apply the random parameter GLM and explainable deep learning to evaluate the heterogenous effects of factors on the taxis' high-range speeding likelihood. Initially, a Beta GLM with random parameters (BGLM-RP) is developed to model the high-range speeding likelihood among taxi drivers. Additionally, XGBoost, a simple convolutional neural network (Simple-CNN), a deeper CNN (DCNN), and a deeper CNN with self-attention (DCNN-SA) are developed. The quantified explanations and illustrations of the factors' heterogenous effects from ML/DL models are derived from pseudo coefficients by decomposing factors' SHapley Additive exPlanations (SHAP) values. All the developed statistical, ML, and DL models are compared in terms of mean absolute errors and mean square errors on testing and full data. Results show that DCNN-SA excels in prediction on testing data, indicating its superior generalization capabilities, while BGLM-RP outperforms other models on full data. The DCNN-SA can reveal the heterogenous effects of factors for both in-sample and out-of-sample data, which is not possible for the random parameter GLM. However, BGLM-RP can reveal larger magnitudes of the factors' heterogenous effects for in-sample data. The signs and significances are identical between the varying coefficients from BGLM-RP and the pseudo coefficients from the ML/DL models, demonstrating the validity and rationale of using the proposed explanation framework to quantify the factors' effects in ML/DL models. The study also discusses the contributions of various factors to the high-range speeding likelihood of taxi drivers.

Real-time accident anticipation for autonomous driving through monocular depth-enhanced 3D modeling

The primary goal of traffic accident anticipation is to foresee potential accidents in real time using dashcam videos, a task that is pivotal for enhancing the safety and reliability of autonomous driving technologies. In this study, we introduce an innovative framework, AccNet, which significantly advances the prediction capabilities beyond the current state-of-the-art 2D-based methods by incorporating monocular depth cues for sophisticated 3D scene modeling. Addressing the prevalent challenge of skewed data distribution in traffic accident datasets, we propose the Binary Adaptive Loss for Early Anticipation (BA-LEA). This novel loss function, together with a multi-task learning strategy, shifts the focus of the predictive model towards the critical moments preceding an accident. We rigorously evaluate the performance of our framework on three benchmark datasets - Dashcam Accident Dataset (DAD), Car Crash Dataset (CCD), and AnAn Accident Detection (A3D), and DADA-2000 Dataset - demonstrating its superior predictive accuracy through key metrics such as Average Precision (AP) and mean Time-To-Accident (mTTA).

Exploring master scenarios for autonomous driving tests from police-reported historical crash data using an adaptive search sampling framework

Crash scenario-based testing is crucial for assessing autonomous driving safety. However, existing studies on scenario generation tend to prioritize concrete scenarios for direct testing, neglecting the construction of fundamentally functional scenarios with a broader range. Police-reported historical crash data is a valuable supplement, yet detecting all potential crash scenarios is laborious. In order to address this issue, this study proposes an adaptive search sampling framework based on deep generative model and surrogate model (SM) to extract master scenario samples from police-reported historical crash data. The framework starts with selecting representative samples from the full crash dataset as initial master scenario samples using various sampling techniques. Evaluation indexes are then constructed, and derived scenario samples are synthesized using the deep generative model. To enhance efficiency, an SM is established to replace the generative model's training and data generation process. Based on the SM, an adaptive search sampling method is developed, which iteratively adjusts the sampling strategy using the Similarity Score to achieve comprehensive sampling. Experimental results demonstrate the notable advantage of the adaptive search sampling method over other sampling methods. Furthermore, statistical analysis and visualization assessments confirm the effectiveness and accuracy of the proposed method.

Real-time crash prediction on express managed lanes of Interstate highway with anomaly detection learning

To facilitate efficient transportation, I-4 Express is constructed separately from general use lanes in metropolitan area to improve mobility and reduce congestion. As this new infrastructure would undoubtedly change the traffic network, there is a need for more understanding of its potential safety impact. Unfortunately, many advanced real-time crash prediction models encounter an important challenge in their applicability due to their demand for a substantial volume of data for direct modeling. To tackle this challenge, we proposed a simple yet effective approach - anomaly detection learning, which formulates model as an anomaly detection problem, solves it through normality feature recognition, and predicts crashes by identifying deviations from the normal state. The proposed approach demonstrates significant improvement in the Area Under the Curve (AUC), sensitivity, and False Alarm Rate (FAR). When juxtaposed with the prevalent direct classification paradigm, our proposed anomaly detection learning (ADL) consistently outperforms in AUC (with an increase of up to 45%), sensitivity (experiencing up to a 45% increase), and FAR (reducing by up to 0.53). The most performance gain is attained through the combination of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) in an ensemble, resulting in a 0.78 AUC, 0.79 sensitivity, and a 0.22 false alarm rate. Furthermore, we analyzed model features with a game-theoretic approach illustrating the most correlated features for accurate prediction, revealing the attention of advanced convolution neural networks to occupancy features. This provided crucial insights into improving crash precaution, the findings from which not only benefit private stakeholders but also extend a promising opportunity for governmental intervention on the express lane. This work could promote express lane with more efficient resource allocation, real-time traffic management optimization, and high-risk area prioritization.

A variable speed limit control approach for freeway tunnels based on the model-based reinforcement learning framework with safety perception

To improve the traffic safety and efficiency of freeway tunnels, this study proposes a novel variable speed limit (VSL) control strategy based on the model-based reinforcement learning framework (MBRL) with safety perception. The MBRL framework is designed by developing a multi-lane cell transmission model for freeway tunnels as an environment model, which is built so that agents can interact with the environment model while interacting with the real environment to improve the sampling efficiency of reinforcement learning. Based on a real-time crash risk prediction model for freeway tunnels that uses random deep and cross networks, the safety perception function inside the MBRL framework is developed. The reinforcement learning components fully account for most current tunnels' application conditions, and the VSL control agent is trained using a deep dyna-Q method. The control process uses a safety trigger mechanism to reduce the likelihood of crashes caused by frequent changes in speed. The efficacy of the proposed VSL strategies is validated through simulation experiments. The results show that the proposed VSL strategies significantly increase traffic safety performance by between 16.00% and 20.00% and traffic efficiency by between 3.00% and 6.50% compared to a fixed speed limit approach. Notably, the proposed strategies outperform traditional VSL strategy based on the traffic flow prediction model in terms of traffic safety and efficiency improvement, and they also outperform the VSL strategy based on model-free reinforcement learning framework when sampling efficiency is considered together. In addition, the proposed strategies with safety triggers are safer than those without safety triggers. These findings demonstrate the potential for MBRL-based VSL strategies to improve traffic safety and efficiency within freeway tunnels.

Advances, challenges, and future research needs in machine learning-based crash prediction models: A systematic review

Accurately modelling crashes, and predicting crash occurrence and associated severities are a prerequisite for devising countermeasures and developing effective road safety management strategies. To this end, crash prediction modelling using machine learning has evolved over two decades. With the advent of big data that provides unprecedented opportunities to better understand the crash mechanism and its determinants, such efforts will likely be accelerated. To gear these efforts, understanding state-of-the-art machine learning-based crash prediction models becomes paramount to summarise the lessons learned from past efforts, which can assist in developing robust and accurate models. This review paper aims to address this gap by systematically reviewing the machine learning studies on crash modelling. Models are reviewed from three aspects of the application: (a) crash occurrence (or real-time crash) prediction, (b) crash frequency prediction, and (c) injury severity prediction. Further, model intricacies that impact model performance are identified and thoroughly reviewed. This comprehensive review highlights specific gaps and future research needs in three aforementioned model applications, such as improper selection of non-crash events for crash occurrence models, the inability of future forecasting of crash frequency models, and inconsistency in injury severity classes. Critical research needs relating to model development, evaluation, and application are also discussed. This review envisages methodological advancements in machine learning models for crash prediction modelling and leveraging big data to better link crashes with its determinants.

Prediction and analysis of likelihood of freeway crash occurrence considering risky driving behavior

The prediction of the likelihood of vehicle crashes constitutes an indispensable component of freeway safety management. Due to data collection limitations, studies have used mainly traffic flow-related variables to develop freeway crash prediction models but rarely have considered the effect of risky driving behavior on the likelihood of crashes. This study employed navigation software to collect driving behavior data and integrated multi-source data that include vehicle speed, traffic volume, and congestion index values. The study also employed the 'synthesizing minority oversampling technique and edited nearest neighbor' (SMOTE + ENN) coupled method for data balance processing. Three freeway crash likelihood prediction models were built based on the binomial logit, eXtreme Gradient Boosting (XGBoost), and support vector machine algorithms, respectively. The Shapley additive explanation (SHAP) algorithm was utilized to explore the effect of each feature variable on the likelihood of crashes. The results show that the prediction accuracy of the XGBoost model is the best of the three compared models. Under the optimal control-to-case ratio (1:1), the prediction accuracy of the XGBoost model reached 0.96 in this study, and the recall rate, specificity, and area-under-the-curve values were 0.86, 0.96, and 0.907, respectively. Comparative test results demonstrate that ranking risky driving behavior into three levels of intensity can effectively enhance the predictive accuracy of the XGBoost model. Moreover, the XGBoost model with its ten-minute time step outperformed the XGBoost model with its five-minute time step in terms of prediction accuracy. The results of the SHAP-based analysis show that the likelihood of highway crashes is high when the traffic congestion level is high and the distribution of the vehicle speed in the upstream roadway section is significant. Also, both sharp acceleration and sharp deceleration lead to greater likelihood of crashes. This paper aims to provide an effective framework for predicting and interpreting the likelihood of freeway crashes, thereby providing guidance for crash prevention, driver training, and the development of traffic regulations.

Crash injury severity prediction considering data imbalance: A Wasserstein generative adversarial network with gradient penalty approach

For each road crash event, it is necessary to predict its injury severity. However, predicting crash injury severity with the imbalanced data frequently results in ineffective classifier. Due to the rarity of severe injuries in road traffic crashes, the crash data is extremely imbalanced among injury severity classes, making it challenging to the training of prediction models. To achieve interclass balance, it is possible to generate certain minority class samples using data augmentation techniques. Aiming to address the imbalance issue of crash injury severity data, this study applies a novel deep learning method, the Wasserstein generative adversarial network with gradient penalty (WGAN-GP), to investigate a massive amount of crash data, which can generate synthetic injury severity data linked to traffic crashes to rebalance the dataset. To evaluate the effectiveness of the WGAN-GP model, we systematically compare performances of various commonly-used sampling techniques (random under-sampling, random over-sampling, synthetic minority over-sampling technique and adaptive synthetic sampling) with respect to dataset balance and crash injury severity prediction. After rebalancing the dataset, this study categorizes the crash injury severity using logistic regression, multilayer perceptron, random forest, AdaBoost and XGBoost. The AUC, specificity and sensitivity are employed as evaluation indicators to compare the prediction performances. Results demonstrate that sampling techniques can considerably improve the prediction performance of minority classes in an imbalanced dataset, and the combination of XGBoost and WGAN-GP performs best with an AUC of 0.794 and a sensitivity of 0.698. Finally, the interpretability of the model is improved by the explainable machine learning technique SHAP (SHapley Additive exPlanation), allowing for a deeper understanding of the effects of each variable on crash injury severity. Findings of this study shed light on the prediction of crash injury severity with data imbalance using data-driven approaches.

Physics-informed multi-step real-time conflict-based vehicle safety prediction

Real-time vehicle safety prediction is critical in roadway safety management as drivers or vehicles can be altered beforehand to take corresponding evasive actions and avoid possible collisions. This study proposes a physics-informed multi-step real-time conflict-based vehicle safety prediction model to enhance roadway safety. Physics insights (i.e., traffic shockwave properties) are combined with data-driven features extracted from deep-learning techniques to improve prediction accuracy. A time series of future vehicle safety indicators are predicted such that vehicles/drivers have enough time to take precautions. The safety indicator at each time stamp is a continuous value that the sign reflects the presence of conflict risks, and the absolute value indicates the conflict risk level to advise different magnitudes of evasive actions. A customized loss function is developed for the proposed prediction model to give more attention to risky events, which are the focus of safety management. The prediction superiority of the proposed model is proven through numerical experiments by comparing it with three benchmarks constructed based on the literature. Further, sensitivity analysis on key model parameters is carried out to advise parameter selections in developing real-world conflict-based vehicle safety prediction applications.

Developing a new real-time traffic safety management framework for urban expressways utilizing reinforcement learning tree

One of the main objectives of an urban traffic control system is to reduce the crash frequency and the loss caused by these crashes on urban expressways. Real-time crash risk prediction (RTCRP) is an essential technique to identify crash precursors so as to take proactive measures to smooth traffic fluctuations. In addition, automatic incident detection (AID) is another important approach to timely detect an incident so as to design countermeasures that reduce any negative impacts on traffic dynamics. With the introduction of disruptive technologies in transport, highly disaggregated large datasets have started to emerge for modelling while existing modelling techniques utilized in RTCRP and AID may not be able to accurately predict traffic crashes in real-time. Therefore, this paper proposes a state-of-the-art reinforcement learning tree (RLT) approach to develop RTCRP model and automatic crash detection (ACD) model similar to AID, and further utilizes it to build a real-time traffic safety management framework for urban expressways with the input of online traffic data streaming. Recorded traffic flow data and historical crash data were extracted and integrated to develop and implement both RTCRP models and ACD models. The prediction results were compared with the frequently used logistic regression (LR), support vector machine (SVM) and deep neural network (DNN) and a sensitivity analysis for variable effects was conducted. The results confirm that RLT outperforms LR, SVM and DNN in developing RTCRP and ACD models. At the cost of 10% false-alarm rate, about 96% of the crashes were predicted or detected correctly by the proposed framework. The results also indicate that: i) collecting more data is helpful to improve the predictive performance and approximatively a minimum sample size of 20 observations per variable is reasonable for training RLT models; and ii) obtaining more factors is beneficial to improve the predictive performance. With the RLT approach, it was demonstrated that selected important variables also have the capability to provide reasonable predictive performance.

Safe route-finding: A review of literature and future directions

While road navigation systems seek to determine the shortest routes between a given set of origin and destination points, there are certain situations in which the fastest route increases the risk of being involved in road crashes. This implies the necessity of integrating safe route-finding into road navigation systems. This study is designed to synthesize the literature on safe route-finding and identify the gaps in the literature for future research. Specifically, a scoping literature review methodology is applied to understand how safety is incorporated in route-finding, even beyond motor vehicle navigation systems. Three databases (Scopus, Web of Science, and IEEE Xplore) are explored, and controlling for inclusion criteria, 40 studies are included in this review. The findings of this review indicated five areas through which safety was considered in route-finding: motor vehicle navigation, public safety, public health, pedestrian and cyclist navigation, and hazardous material transportation. The measurement of safety was found challenging with inconsistencies in safety quantification approaches. The safe route-finding algorithms were investigated based on their predictive/reactive, static/dynamic, and centralized/decentralized characteristics. Based on the critical review of the safe route-finding algorithms, availability of real-time data sources, accurate real-time and disaggregated crash risk prediction models, trade-off between time and safety in road navigation tools, and centralized safe route-finding are highlighted as the requirements and challenges in considering safety in road navigation systems. This study outlines a research agenda to address the identified challenges in safe route-finding.

A data-centric weak supervised learning for highway traffic incident detection

Using the data from loop detector sensors for near-real-time detection of traffic incidents on highways is crucial to averting major traffic congestion. While recent supervised machine learning methods offer solutions to incident detection by leveraging human-labeled incident data, the false alarm rate is often too high to be used in practice. Specifically, the inconsistency in the human labeling of the incidents significantly affects the performance of supervised learning models. To that end, we focus on a data-centric approach to improve the accuracy and reduce the false alarm rate of traffic incident detection on highways. We develop a weak supervised learning workflow to generate high-quality training labels for the incident data without the ground truth labels, and we use those generated labels in the supervised learning setup for final detection. This approach comprises three stages. First, we introduce a data preprocessing and curation pipeline that processes traffic sensor data to generate high-quality training data through leveraging labeling functions, which can be domain knowledge-related or simple heuristic rules. Second, we evaluate the training data generated by weak supervision using three supervised learning models-random forest, k-nearest neighbors, and a support vector machine ensemble-and long short-term memory classifiers. The results show that the accuracy of all of the models improves significantly after using the training data generated by weak supervision. Third, we develop an online real-time incident detection approach that leverages the model ensemble and the uncertainty quantification while detecting incidents. Overall, we show that our proposed weak supervised learning workflow achieves a high incident detection rate (0.90) and low false alarm rate (0.08).

Estimation of the Origin-Destination Matrix for Trucks That Use Highways: A Case Study in Chile

Nowadays, freight transport is crucial in the functioning of cities worldwide. To dig further into the understanding of urban freight transport movements, in this research, we conducted a case study in which we estimated an origin-destination matrix for the trucks traveling on Autopista Central, one of Santiago de Chile’s most important urban highways. To do so, we used full real-world vehicle-by-vehicle information of freight vehicles’ movements along the highway. This data was collected from several toll collection gates equipped with free-flow and automatic vehicle identification technology. However, this data did not include any vehicle information before or after using the highway. To estimate the origins and destinations of these trips, we proposed a multisource methodology that used GPS information provided by SimpliRoute, a Chilean routing company. Nevertheless, this GPS data involved only a small subset of trucks that used the highway. In order to reduce the bias, we built a decision tree model for estimating the trips’ origin, whose input data was complemented by other public databases. Furthermore, we computed trip destinations using proportionality factors obtained from SimpliRoute data. Our results showed that most of the estimated origins belonged to outskirt municipalities, while the estimated destinations were mainly located in the downtown area. Our findings might help improve freight transport comprehension in the city, enabling the implementation of focused transport policies and investments to help mitigate negative externalities, such as congestion and pollution.

Transfer learning for spatio-temporal transferability of real-time crash prediction models

Real-time crash prediction is a heavily studied area given their potential applications in proactive traffic safety management in which a plethora of statistical and machine learning (ML) models have been developed to predict traffic crashes in real-time. However, one of the fundamental issues relating to the application of these models is spatio-temporal transferability. The present paper attempts to address this gap of knowledge by combining Generative Adversarial Network (GAN) and transfer learning to examine the transferability of real-time crash prediction models under an extremely imbalanced data setting. Initially, a baseline model was developed using Deep Neural Network (DNN) with crash and microscopic traffic data collected from M1 Motorway in the UK in 2017. The dataset utilised in the baseline model is naturally imbalanced with 257 crash cases and 16,359,163 non-crash cases. To overcome data imbalance issue, Wasserstein GAN (WGAN) was utilised to generate synthetic crash data. Non-crash data were randomly undersampled due to computational limitations. The calibrated model was then applied to predict traffic crashes for five other datasets obtained from M1 (2018), M4 (2017 & 2018 separately) and M6 Motorway (2017 & 2018 separately) by using transfer learning. Model transferability was compared with standalone models and direct transfer from the baseline model. The study revealed that direct transfer is not feasible. However, models become transferable temporally, spatially, and spatio-temporally if transfer learning is applied. The predictability of the transferred models outperformed existing studies by achieving high Area Under Curve (AUC) values ranging between 0.69 and 0.95. The best transferred model can predict nearly 95% crashes with only a 5% false alarm rate by tuning thresholds. Furthermore, the performances of transferred models are on par with or better than the standalone model. The findings of this study proves that transfer learning can improve model transferability under extremely imbalanced settings which helps traffic engineers in developing highly transferable models in future.