Content-aware vision-based vehicle tracking for frame-skipped videos

• Main Author: Cempron, Jonathan Paul C.
• Format: text
• Language: English
• Published: Animo Repository 2021
• Subjects: Computer vision; Streaming video; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etdm_comsci/3
• Institution: De La Salle University

Vision-based object tracking aims to approximate the trajectory of an object as it is observed to move in a video. Mainstream algorithms for this task assume that the object’s motion is smooth and require high frame rates. However, camera networks that stream video exhibit frame skipping, which is a visual degradation that causes object motion to be disjointed. This work addressed the frame skipping problem on the object tracking algorithm level by leveraging content information and a dynamic selection of motion and appearance features. A motion-based tracking algorithm is used if the video possesses a high frame rate and an appearance-based tracking algorithm is used if the frame rate is low. The performance of the tracking algorithm is assessed using ClearMOT metrics. The tracking algorithm developed in this work outperformed the classical tracking algorithm.