Improving Multi-Object Tracking-by-Detection in Video via Fusion of Kalman Filter and Deep Learning

Bani Hashemi, Seyed Mojtaba; Zoghi, Mohammad Reza

doi:10.22034/joeds.2026.555293.1105

Improving Multi-Object Tracking-by-Detection in Video via Fusion of Kalman Filter and Deep Learning

Document Type : Original Article

Authors

Seyed Mojtaba bani Hashemi ¹

Mohammad Reza Zoghi ²

¹ the faculty of Electrical and Computer Engineering, university of kashan, kashan, iran

² The Faculty of Electrical and Computer Engineering, university of kashan, kashan, iran

10.22034/joeds.2026.555293.1105

Abstract

Multi-object tracking is a fundamental computer-vision task that has drawn ever-increasing attention because of its scientific and commercial potential. Nevertheless, accurate object tracking remains highly challenging; these challenges include the high similarity and density of detected objects. Moreover, occlusion and viewpoint changes can occur as objects move. In this paper, a framework for real-time multi-object tracking is introduced that is based on a modified version of the SORT algorithm. Multi-object tracking is divided into two parts. In the first part, object detection is performed using the YOLO family; if information is lost at this stage, compensating for this lost information later is impossible. The second part concerns object tracking, which itself comprises three stages: first, feature extraction, for which transfer learning with the YOLOv8 family is used; second, position prediction using the Kalman filter; and third, data association and object matching, for which the Hungarian algorithm is employed. In the data-association stage, the use of deep-learning methods has recently expanded. Finally, the MOTA metric was adopted as the result, yielding 65.3, which is 7.2 % better than the reference paper.

Keywords

Multi Object Tracking

Tracking by Detection

Data association

Deep Learning

Image Processing