EF-StrongSORT: An Enhanced Feature StrongSORT Model for Multi-Object Tracking

Abstract

Multi-object tracking (MOT) faces persistent challenges owing to the complexities introduced by occlusions, dynamic appearance variations, and the rapid motion of objects within a scene. These issues are further complicated by the need for robust identity management and consistent object re-identification across frames. To improve the performance of multi-object tracking, this study introduces EF-StrongSORT, which extends the StrongSORT model, incorporating advanced object detection, efficient feature extraction, and identity management techniques. The EF-StrongSORT demonstrates an improvement over conventional tracking methods by achieving higher accuracy and robustness in challenging scenarios. The experimental results show that the EF-StrongSORT enhances the performance of multi-object tracking techniques and is better than existing approaches on the MOT17, MOT20 and DanceTrack benchmarks. On the MOT17 dataset, EF-StrongSORT outperformed StrongSORT with improvements of +5.5 in HOTA, +7.7 in MOTA, +4 in IDF1, and a decrease of 828 in IDS. On the MOT20 dataset, EF-StrongSORT showed improvements of +2 in HOTA, +6.3 in MOTA, +1.1 in IDF1, and a reduction of 16 in IDS compared to StrongSORT. On the DanceTrack dataset, EF-StrongSORT achieved improvements of +4.3 in HOTA, +2.2 in MOTA, and +1.4 in AssA compared to the latest state-of-the-art model, LQTTrack. These results emphasize the contributions of the proposed model to the improvement of quality and efficiency of multi-object tracking systems targeted for specific problems, including object appearance changes and occlusions.