Abstract: Vehicle trajectory extraction on urban roads is crucial for intelligent transportation supervision, but existing techniques suffer from low detection accuracy and broken trajectories due to identity hopping. In this paper, a composite framework that fuses improved YOLOv7-tiny detection, StrongSORT tracking and Savitzky-Golay filter optimization is proposed. The framework is capable of efficiently extracting the trajectories of different vehicle targets using urban road surveillance video data collected by traffic monitoring devices. Based on experimental evaluation, the IYSSG framework performs well in three main tasks. In vehicle detection, the improved YOLOv7-tiny algorithm ensures the detection speed, while precision, recall rate and mAP@0.5 increase by 2.5%, 8.5%, and 3.7%, respectively, compared with the original YOLOv7-tiny algorithm. In terms of vehicle tracking, the StrongSORT algorithm achieves a 4.92% and 2.7% improvement in MOTA and MOTP metrics, respectively, compared with the DeepSORT algorithm. In terms of vehicle trajectory extraction and optimization, the Savitzky-Golay filtering algorithm effectively solves the problems of missing trajectory points and unsmooth trajectory due to objective factors such as video jitter and algorithmic errors, which helps the researchers to extract accurate vehicle trajectories from the traffic surveillance video for better analysis and localization of traffic problems.

Key words: YOLOv7-tiny, target detection, deep learning, multi-target tracking, trajectory extraction, urban road, vehicle track