Deep learning–based multi-target detection and flow prediction in complex traffic systems supported by the CV (chan–vese) segmentation model

Abstract In recent years, accurate multi-target detection and vehicle flow prediction have become essential for intelligent transportation management in complex traffic environments. However, conventional detection-based methods often exhibit limited robustness to occlusions, illumination variations, and re-identification (Re-ID) challenges in dynamic scenes. To overcome these limitations, this study introduces a novel integrated deep learning framework that leverages the Chan-Vese (CV) segmentation model for occlusion-resilient image preprocessing, refining target contours and generating high-quality datasets to bolster subsequent analysis. Building on this foundation, the framework combines YOLOv3 for precise object detection, a residual-enhanced Re-ID with split-transform-merge structures for efficient feature association, and Kalman filtering for robust trajectory estimation and state prediction. Trained and validated on real surveillance videos captured from multiple expressway service areas, the proposed system achieves superior performance. It attains Multiple Object Tracking Accuracy (MOTA) and Multiple Object Tracking Precision (MOTP) scores of 50.56% and 70.56%, respectively, while also ensuring enhanced identity consistency and reduced false tracks. These scenario-based datasets reflect practical traffic conditions such as varying vehicle densities, lighting, and occlusions, ensuring that the model’s optimization directly corresponds to real-world highway monitoring and management needs. Comparative evaluations against alternative detection and correlation algorithms confirm its effectiveness in multi-target detection and flow prediction, enabling real-time applications such as congestion alerts, dynamic resource allocation, and optimized routing to enhance traffic safety and efficiency.