Vehicle Path Tracing of Traffic Congestion Points and Sections on Urban Expressways Based on Deep Learning

Abstract:

This paper aims to overcome the limitations of existing research that simplifies the traffic congestion source-tracing problem into path flow estimation or congestion correlation analysis. It proposes a more comprehensive and effective system for tracing vehicle paths in traffic-congested sections of urban expressways. Using the path as the basic analysis unit， it develops an innovative unified framework integrating both path flow estimation and congestion correlation analysis. Additionally， it proposes a method based on the route-based deformable convolution long short-term memory neural network （RSDC-LSTM）. The model consists of three core modules： constructing a path state feature set based on historical path flow data and short-term prediction data； quantifying the dynamic influence weights of each path on traffic congestion through a collaborative modeling of the multi-path convolutional long short-term memory network and the soft-attention mechanism； and using the deformable convolutional neural network to capture the spatial-topological correlation features of congested sections and achieve the evaluation of path importance in both spatial and temporal dimensions. Empirical research shows that RSDC-LSTM effectively identifies key paths of traffic congestion and ranks their influence. By regulating the top 10% of high-influence paths， peak travel speeds can be increased by 23.36%， while the number of stops and delay time can be reduced by up to 29.41% and 43.82% respectively. The RSDC-LSTM method proposed provides a quantifiable decision-making framework for developing dynamic traffic control strategies and contributes to improving the traffic operation efficiency of urban expressways.