A LSTM-BN Based Model for Real-time Vehicle Collision Risk Prediction

The development of vehicle-to-infrastructure cooperation (V2I) technology enables real-time dynamic information interaction between humans, vehicles, and roads, providing technical support for real-time vehicle collision risk prediction. Current models for predicting vehicle collision risk have limitations in capturing dynamic features and performing uncertainty inference. Thus, a two-layer progressive model based on long short-term memory-Bayesian network (LSTM-BN) is proposed. The LSTM-BN model utilizes LSTM at the bottom layer to capture the temporal evolution characteristics of vehicle states by setting a 0.1 s sampling window. At the top layer, the model performs probabilistic inference of collision risk through the BN. Based on data from naturalistic driving experiments in Wuhan, the time-to-collision (TTC) index is used to define and quantify the collision risk of vehicles (cars). This process yields 25, 826, 19, 344, and 4, 051 frames of no-risk, low-risk, and high-risk data, respectively. After preliminarily determining the variables affecting collision risk, the decision tree method is employed to extract key causal features from the dimensions of driver-vehicle-road. Interaction indicators, such as driver proficiency, speed difference, road type, and headway, are further excavated and characterized to reflect the non-linear coupling relationships among multi-dimensional factors. A Bayesian network reflecting these coupling relationship indicators is then constructed to evaluate the current vehicle collision risk. On this basis, to compensate for the deficiency of static probabilistic models in sequential prediction, LSTM is used to predict vehicle state variables. These variables are then mapped into the BN to infer the collision probability at the next time step. Through the above steps, the model achieves a precise prediction of real-time vehicle collision risk under dynamic evolution. To verify the effectiveness of the proposed model, its prediction performance is evaluated from four aspects: robustness, accuracy, timeliness, and complexity. Results from comparative modeling indicate that the LSTM-BN model achieves the highest accuracy at 91%. This performance is 7% and 12% higher than the accuracies of the support vector machine (SVM) and random forest (RF) models, respectively. Furthermore, its index fluctuations across different sampling periods are all below 0.1, demonstrating strong robustness.