Trajectory Similarity-Based Traffic Flow Analysis Using YOLO+ByteTrack

초록

The proliferation of vehicles in modern society has led to increased traffic congestion and accidents, necessitating advanced traffic monitor-ing systems. Nevertheless, current systems encounter challenges in balancing effective vehicle tracking with privacy protection and face diffi-culties in anomaly detection across diverse traffic environments. This study introduces an innovative approach to traffic flow analysis using deep learning-based vehicle trajectory similarity comparison. The objectives are to develop a real-time vehicle detection and tracking system that protects privacy and to identify anomalous traffic flows through trajectory similarity-based grouping. The methodology employs a pipeline combining YOLO models for object detection, ByteTrack for vehicle tracking, and trajectory similarity metrics for grouping and analysis. Experiments were conducted using high-quality CCTV traffic video datasets from AI-Hub, evaluating various YOLO models and tracking performance. The YOLOv7x model exhibited the best performance with a mAP@0.5 of 0.708 and 87.9843 FPS, making it suitable for real-time vehicle detection. When combined with ByteTrack, the system achieved a MOTA of 0.289, a MOTP of 0.837 and an IDF1 of 0.725, indicating stable tracking performance. Trajectory similarities were analyzed using Cosine Similarity, Jensen-Shannon Divergence, and Euclid-ean Distance Similarity, enabling comprehensive evaluation of vehicle movements. The proposed method minimizes privacy concerns while enabling effective real-time traffic flow analysis. However, limitations include experiments conducted only in limited road environments and lack of performance verification under extreme conditions. Future research will focus on expanding the dataset, improving model performance through GANs, and integrating anomaly detection algorithms to enhance the system’s capability in managing diverse traffic scenarios.

키워드