PASTGCN: A Real-Time Intersection Conflict Risk Prediction Model via Spatiotemporal Graph Convolutional Networks with Multimodal Sensor Fusion

Abstract

With the intelligent upgrade of urban transportation systems, intersections, as key nodes in traffic flow, have significant importance in predicting conflict risks in real-time and accurately for proactive prevention and control, as well as improving traffic efficiency. Traditional methods mainly rely on historical statistical patterns or single sensor data, which makes it difficult to effectively capture the dynamic spatiotemporal correlation characteristics of intersection traffic flow, resulting in insufficient timeliness and accuracy of predictions. To address this issue, this study proposes a real-time prediction model for PASTGCN intersection conflict risk based on Spatiotemporal Graph Convolutional Network (STGCN). The architecture includes five core units: temporal convolution, spatial convolution, graph structure learning, spatiotemporal position embedding, and output. The spatiotemporal position embedding unit generates a dynamic adjacency matrix with spatiotemporal heterogeneity through two sets of learnable node embedding vectors and temporal position vectors, combined with self attention mechanism. The graph structure learning unit adapts to the time-varying characteristics of traffic flow through a dual layer mechanism of "static topology+dynamic adjustment" (the static graph learning submodule uses Graph WaveNet to construct the basic adjacency matrix through matrix decomposition, and the time aware graph learning submodule integrates dynamic optimization of time information). The model integrates multi-source data from video detection, radar sensing, and geomagnetic sensors to construct a dynamic heterogeneous graph that includes node attributes, edge weights, and global road network topology. An improved graph convolutional layer is used to extract spatially relevant features, and a time gated recurrent unit (GRU) is used to determine the collision probability of traffic flow in all directions within 30 seconds. The experiment is based on traffic trajectory data validation of 100 typical intersections in a mega city from January to June 2024. The results showed that the Mean absolute error (MAE) of the model in the binary classification task of conflict events was 0.12, the root mean square error (RMSE) was 0.18, and the F1 score reached 89.2%, which was significantly optimized compared to traditional LSTM models and static GCN models; Real time prediction delay is stable within 180ms, which can meet the millisecond level response requirements of intersection signal control systems. This study provides high-precision and low latency technical support for proactive prevention and management of urban traffic conflicts.