2.2. Research Outcomes

Analysis of Roadway Safety. Tobin et al. ^[2] highlighted that the relative crash risk is significantly higher during periods of precipitation compared to non-precipitation periods. In a similar vein, Wen et al. ^[9] demonstrated that the importance of risk factors varies across different crash types. For rear-end (RE) crashes, the speed limit emerged as a more crucial risk factor than lane width, right/left shoulder width, and median width. Conversely, for run-off-road (ROR) crashes, the opposite relationship was observed. 

Analysis of Freeway Safety. Several research findings have contributed to our understanding of factors influencing crash occurrence and risk mitigation. Rahman et al. ^[3] identified a high variation in speed at a downstream station and high traffic volume at an upstream station as factors increasing the likelihood of crash occurrence. Zheng et al. ^[10] further supported this by demonstrating that the downstream average speed was the best crash precursor variable across different segment types. The effectiveness of warning information systems in preventing secondary crash risks was shown by Son et al. ^[8], with Jang et al. ^[11] reporting a significant reduction in the crash potential through the provision of warning information. 

Analysis of Intersection Safety. Several research findings have shed light on various aspects related to crash risk and intersection safety. Kwon et al. ^[12] emphasized the significance of intersection characteristics, such as the proportional area of sky and roadway, in influencing the perceived crash risk among school-aged children. Mitra et al. ^[13] identified multiple factors that significantly influenced both the frequency and severity of crashes, including blocked carriageways, approach traffic volume, traffic configuration, type of minor road, presence of protected right turning phase, tram stops, all-red time, visibility of road markings, and non-motorized traffic. 

Analysis of Vehicle, Pedestrian, and Cyclist Safety. Research findings in the field of road safety have provided insights into various factors affecting the crash risk for different road users. For vehicle-related factors, Baikejuli et al. ^[14] highlighted the contribution of multifactor interaction, such as environmental and vehicular factors, in increasing the crash probability for heavy-truck fatal crashes. 

3. Operation or Control

3.1. Method

Convolutional Neural Networks (CNNs). Convolutional Neural Networks (CNNs) are deep learning models specifically designed for analyzing visual data like images. They use layers of filters to extract meaningful features from the input images. These filters perform convolution operations, highlighting patterns and structures. The network then learns to recognize complex features by stacking multiple convolutional layers. Pooling layers downsample the feature maps to capture important information. Fully connected layers process the extracted features and make predictions. CNNs are trained using labeled data, optimizing their parameters to minimize prediction errors. CNNs have revolutionized computer vision tasks by automatically learning relevant features directly from images, enabling them to achieve high accuracy in tasks like image classification and object detection. 

LSTM. LSTM (Long Short-Term Memory) is a type of neural network that excels at processing sequential data, like text or time series. Unlike traditional recurrent neural networks (RNNs), LSTMs can remember long-term dependencies in the data by selectively retaining or discarding information using memory cells and specialized gates. This allows them to capture patterns and make accurate predictions in tasks involving sequences. Previous research often combines the convolutional layers with the long short-term memory (LSTM) neural network in a unified deep learning framework, utilizing LSTM to extract temporal features.

XGBoost. XGBoost, short for Extreme Gradient Boosting, is a powerful machine learning algorithm known for its exceptional performance in various predictive modeling tasks. It belongs to the gradient boosting family of algorithms and combines the strengths of decision trees with a boosting approach. XGBoost iteratively builds an ensemble of weak decision tree models, optimizing a specific objective function by minimizing the residuals of the previous model. It incorporates regularization techniques to prevent overfitting and employs parallel processing to accelerate training. XGBoost is highly efficient, scalable, and capable of handling large-scale datasets, making it a popular choice for tasks such as regression, classification, and ranking. 

Reinforcement Learning (RL). Reinforcement learning is a branch of machine learning, where an agent learns to make sequential decisions in an environment to maximize a cumulative reward signal. It involves an agent interacting with an environment, taking actions, receiving feedback in the form of rewards, and adjusting its behavior over time through trial and error. Reinforcement learning is applicable to traffic management because it can optimize traffic flow, reduce congestion, and improve overall efficiency. By treating traffic management as a sequential decision-making problem, reinforcement learning algorithms can learn to control traffic signals, dynamically adjust the traffic flow, and optimize the traffic patterns based on real-time feedback.

Generative Adversarial Network. GAN stands for Generative Adversarial Network, which is a type of deep learning model consisting of two main components: a generator and a discriminator. The generator is trained to produce synthetic data samples that exhibit similarity to the real data, while the discriminator is trained to differentiate between the real and synthetic data. They are trained simultaneously in a competitive manner, with the goal of the generator producing data that can fool the discriminator.

Graph Neural Network. Graph Neural Networks (GNNs) are a type of neural network specifically designed to handle data with graph structures. GNNs can effectively capture and model the relationships between entities represented as nodes and edges in a graph. In the context of traffic management, GNNs can be applied to various safety-related tasks. For instance, GNNs can analyze road networks and capture the complex dependencies between different road segments, traffic intersections, and their associated attributes (e.g., traffic volume and speed limits). By leveraging this information, GNNs can predict traffic congestion, identify accident-prone areas, and even optimize traffic signal timings for improved safety. GNNs enable a holistic view of the traffic system by considering the spatial relationships and interactions between road elements, leading to more accurate and context-aware safety predictions and interventions in traffic management.

Transformer. The Transformer model is a powerful neural network architecture primarily used for sequence-to-sequence tasks, such as machine translation and natural language processing. Unlike traditional recurrent neural networks (RNNs), Transformers rely on self-attention mechanisms, enabling them to capture global dependencies between input elements efficiently. The model comprises an encoder and a decoder, each consisting of multiple layers of self-attention and feed-forward neural networks. The Transformer’s attention mechanism allows it to attend to relevant parts of the input sequence, facilitating parallel processing and capturing long-range dependencies effectively.

3.2. Research Outcomes

Control Methods for Highway, Roadway, and Urban Arterials. In general, researchers have explored strategies for enhancing deep learning models to attain improved performance, while simultaneously addressing the challenge of data imbalance and devising optimization techniques for it. Less complex deep models were found to achieve better performance according to Huang et al. ^[15], while Li et al. ^[16] demonstrated that their LSTM-CNN model achieved superior performance in terms of the AUC value, false alarm rate, and sensitivity compared to other models. 

Control Methods for Intersections. Researchers have made notable contributions in optimizing safety and efficiency at signalized intersections. Ghoul et al. ^[17] proposed a signal-vehicle coupled control system that effectively enhanced safety with low computational intensity. Hu et al. ^[18] demonstrated that deep learning models, particularly the CNN model with an accuracy of 93.8%, are recommended for predicting risk levels at intersections by combining CV data and deep learning networks. Even with low CV penetration rates, this approach showed promise in determining crash risks.

Control Methods for Road Regions. In recent studies, several approaches have been proposed to address different aspects of traffic risk prediction. Zhou et al. ^[19] devised a dynamic graph neural network to capture real-time traffic variations and correlations among subregions, showcasing the efficacy of multitask learning. Wang et al. ^[20] combined a spatial–temporal geographical module with Graph Convolutional Networks (GCN) and Gated Recurrent Units (GRU), showcasing improved performance. 

4. Crash Risk Prediction

Zheng et al. ^[21] introduced a method for forecasting the real-time probability of crashes at signalized intersections, focusing on individual signal cycles. Their approach relied on extracting traffic conflicts from informative vehicle trajectories as a foundation for crash prediction. To address the challenges posed by nonstationarity and unobserved heterogeneity in their model, they established a Bayesian hierarchical structure. The main contribution of their research lay in their novel measurement of traffic conflicts. Specifically, they employed computer vision techniques to extract traffic conflicts, quantified as modified time to collision, along with three cycle-level traffic parameters (shock wave area, traffic volume, and platoon ratio) from video data. In a subsequent study by Gu et al. ^[22], they further explored the realm of intersection safety by highlighting the advantages of connected vehicle technology. This innovation offers abundant vehicle motion data, establishing a stronger link between crash occurrence and driving behaviors. The authors also addressed the challenge of spatial dependence in crash frequency and the multitude of driving features involved in the prediction process. The novelty of their research lay in the introduction of a new artificial intelligence technique known as Geographical Random Forest (GRF). This technique effectively handled spatial heterogeneity and incorporated all the potential predictors. 

In addition to crash risk prediction on highways, class imbalances are also present in the prediction of driving safety risks. To address this issue, Chen et al. ^[23] introduced a novel approach using a deep autoencoder network with L1/L2-nonnegativity constraints and cost sensitivity. This method effectively handled class imbalances and enhanced the prediction performance by determining the optimal sliding window size and automatically extracting hidden features from driving behaviors. The limitation of previous studies in analyzing driver factors and driving maneuvers, due to the absence of disaggregated driving or accident data, was highlighted by Mahajan et al. ^[24]. To overcome this issue, they introduced a method that considered both the likelihood and potential severity of a collision. This study focused on estimating the rear-end crash risk in specific traffic states and emphasized the significance of comprehending the evolution of crash risk under diverse traffic conditions for real-time crash prediction. Taking advantage of the progress made in deep learning technology, Li et al. ^[25] proposed an attention-based LSTM model for predicting lane change behavior. Their objective was to enhance both the accuracy and interpretability. Their approach involved two components: a prejudgment model utilizing a C4.5 decision tree and bagging ensemble learning and an LSTM model with an attention mechanism for multistep lane change prediction.

A rear-end collision prediction method for smart cities was proposed by Wang et al. ^[26]. The CNN model was used for prediction using real trajectory data, while synthetic oversampling using the genetic theory of inheritance was employed to address the class imbalance. Trirat et al. ^[27] introduced a deep gusion network for predicting traffic accident risk across urban areas, integrating hazardous driving statistics collected from in-vehicle sensors. The study examined the correlation between dangerous driving offenses and traffic accidents, revealing a strong correlation in terms of the location and time. The work by Elassad et al. ^[28] addressed the importance of real-time crash prediction and the development of fusion frameworks for intelligent transportation systems. They explored the use of machine learning models and fusion techniques to improve crash predictions by considering diverse data sources, including driver inputs, physiological signals, vehicle kinematics, and weather conditions. The paper highlighted the significance of addressing the class imbalance and presented the effectiveness of boosting in combination with k-NN, Naïve Bayes, Bayesian networks, and SVM with MLP as the meta-classifier.