Abstract:
Since information as the demand for automated scene understanding and
persistent visual monitoring increased, Multi-Object Tracking (MOT) has become an
essential component in various real-life applications, like smart video surveillance,
traffic monitoring, robotics, sports analytics, customer behavior analysis and drone
control. Nevertheless, MOT remains challenging due to complex visual conditions such
as heavy occlusion, frequent identity switching, variable illumination, and inter-object
similarity. The proposed system outperforms a novel MOT framework integrating the
YOLOv9 multi-object detector with a DeepSORT algorithm based on the Generalized
Intersection over Union (GIoU) metric and an enhanced matched and filtering
algorithm to improve detection accuracy, reduce ID switching and occlusion
challenges. This integration aims to effectively tackle the critical challenges of ID
switching and occlusion, thereby enhancing multi-object tracking accuracy on the MOT
16 datasets. The suggested multi-object tracking system applies a refined matched filter
approach featuring two major improvements: a class-based validation step ensuring
detection consistency and a re-identification memory component preserving unmatched
tracks for possible restoration. The proposed system results are compared with the
DeepSORT algorithm, which uses standard Intersection over Union (IoU).
Measurement results show the proposed conversion method greatly improves tracking
performance, offering better accuracy and identity preservation than the standard IoU
system. The experiment demonstrated that the improved system keeps object identities
consistent during occlusions, lowers ID switches, and boosts overall tracking accuracy
over the DeepSORT baseline.