Observation-Centric with Appearance Metric for Computer Vision-Based Vehicle Counting

Abstract

Tracking objects in video sequences is a key step for applications involving computer vision like traffic monitoring and security systems. Occlusion is a frequent problem in object tracking which can result in the tracker losing track of the occluded object or misidentifying it with the occluding object. Moreover, the limited memory and computing power of traffic analysis systems presents a scaling problem, especially in object tracking applications. This paper aims to improve object tracking performance by minimizing data association errors in low frame rate tracking applications. Reducing frame rates alleviates memory and computing power limitations, and utilizing a tracker that can handle occlusion can address occlusion-related issues in object tracking. The proposed tracking method, Mask-OCSORT, uses the observation-tracking method with cosine similarity, intersection-over-union, and velocity consistency metrics for the association problem. The paper analyzes the effect of using bounding box and mask predictions of deep learning models in generating tracks. This study uses evaluation metrics like HOTA, MOTA, and IDF1 to assess the proposed tracking method, and employs evaluation metrics such as precision, recall, and F-score to assess the counting based on generated IDs from the tracking method. The study applied the Mask-OCSORT tracking for vehicle counting application and achieved an F-score of 87.18% at 5 frames per second (fps), and 75% at 1 fps.