Urban traffic flow analysis via cross-scale cross-attention mechanisms

With the continuous advancement of urbanization, the growing ownership of private vehicles has exacerbated road traffic challenges, thereby highlighting the critical importance of enhancing urban traffic flow analysis. In this study, we propose an innovative urban traffic flow analysis model that integrates cross-scale information through deep neural networks and cross-attention mechanisms. Specifically, vehicle detection is conducted using the YOLO algorithm on video data, followed by homography-based matching to obtain traffic indicators such as flow and density, and their temporal variations. Furthermore, a DTW-LSTM model is employed to predict traffic conditions, thereby providing decision support for urban traffic management. Experimental validation demonstrates the robustness and stability of the proposed model, achieving MAE = 1.82, RMSE = 1.70, and MAPE = 0.23. These results indicate that the overall error remains within an acceptable range, confirming the strong performance of the model.