Research on Urban Traffic Flow Statistics and Prediction Models Based on Machine Vision

With the continuous advancement of urbanization, traffic congestion has become a critical issue restricting urban development and impacting residents' travel efficiency. The core of building an Intelligent Transportation System (ITS) lies in achieving accurate perception and prediction of traffic flow. To address the limitations of traditional detection methods, such as high deployment costs and difficult maintenance, this study proposes and implements a comprehensive machine vision solution based on deep learning. The system first designs an improved YOLOv9 vehicle detection model. By integrating the SimAM parameter-free attention mechanism and adding a P2 small-object detection layer, it significantly enhances detection accuracy under challenges like complex weather, varying illumination, and vehicle occlusion, achieving a mean Average Precision (mAP@0.5) of 92.7% on the UA-DETRAC dataset. Furthermore, the DeepSORT multi-object tracking algorithm is optimized by adopting a vehicle-specific Re-ID model and a bidirectional virtual detection line counting logic, increasing traffic flow counting accuracy to over 97% and reducing identity switches (IDSW) by nearly 60%. Experimental results demonstrate that the system achieves high-precision, highly robust real-time traffic flow statistics, providing a reliable data foundation for subsequent traffic flow prediction and management decisions, with significant theoretical value and broad application prospects.