A framework for object tracking in videos with complex backgrounds and obstructions

作     者：Chang, Tsui-Ping Chen, Tzer-Long Hsiao, Tsung-Chih 

作者机构：Department of Computer Science and Information Engineering National Taichung University of Science and Technology Taichung Taiwan Department of Healthcare Administration and Medical Informatics Kaohsiung Medical University Gaoxiong807 Taiwan Department of Radio Television and Film Shih Hsin University Taipei11604 Taiwan 

出 版 物：《Neural Computing and Applications》 (Neural Comput. Appl.)

年 卷 期：2025年

页      面：1-20页

核心收录：

学科分类：1201[管理学-管理科学与工程(可授管理学、工学学位)] 08[工学] 0813[工学-建筑学] 0803[工学-光学工程] 0814[工学-土木工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Funding was provided by National Science and Technology Council (Grant No. 113-2635-E-025 -001 -) 

主　　题：Image enhancement 

摘      要：The research of object tracking in videos utilizes computer vision and machine learning techniques to identify and track objects in the consecutive image frames of videos. The popular algorithms used in the research are YOLO, DeepSORT, StrongSORT, and ByteTrack. YOLO has gained recognition for its hardware efficiency and impressive performance in object detection tasks. It efficiently detects objects in the consecutive image frames of videos. On the other hand, DeepSORT, StrongSORT, and ByteTrack leverage the prediction capabilities to estimate the positions of objects in the next image frames. These algorithms play a significant role to enhance the capabilities of object tracking systems. Recently, many researchers have proposed their methods to enhance the accuracy in object tracking tasks. However, the division of videos into the consecutive image frames for tracking objects can pose challenges, especially when the videos have complex backgrounds and many obstructions in the image frames. In fact, the complex backgrounds with many obstructions in image frames decrease the accuracy of tracking objects in applications. In this study, we present TraObs, a framework for object Tracking in complex backgrounds with Obstructions in the image frames of videos. In TraObs, two novel mechanisms CFM and M-controller are proposed for detectors and trackers to enhance the accuracy of object tracking performance. CFM addresses the tracked objects which are missing in some image frames of videos, while M-controller makes sure the tracked objects with the correct object ids in the system. Some experiments with four testing datasets and benchmark datasets are designed and used respectively to analyze the performance of TraObs. Furthermore, the HOTA and MOTA metrics are used to evaluate our proposed methods. The experimental results also prove the effectiveness of CFM and M-controller in our framework TraObs. © The Author(s), under exclusive licence to Springer-Verlag London Ltd.,