Authors - Phway Phway Aung, Tin Zar Thaw Abstract - Multi-object tracking (MOT) is widely applied in surveillance, traffic monitoring, and autonomous systems. Most MOT systems are created by combining DeepSORT and YOLO series. The original DeepSORT relies on IoU as-sociation-based matching and a fixed age threshold deletion algorithm which of-ten leads to incorrect associations, premature track removal, and frequent ID switches under occlusion or fast motion. To address these limitations, YOLOv9-Based Multi-Object Tracking System is proposed by using GIoU for more reliable geometric association and the enhanced filtering algorithm that are integrated class validation, motion uncertainty estimation with consign similarity, and a Re-Identification (ReID) memory buffer for reducing ID switching. To analyze the performance of the proposed MOT system we compare two cases: IoU and GIoU on Original DeepSORT and the improved DeepSORT and original DeepSORT based on MOT16 videos’ sequences. Experimental evaluation demonstrates that the proposed YOLOv9-Based Multi-Object Tracking System achieves more sta-ble and accurate tracking performance compared to the original DeepSORT system in two cases.
Open Zoom