B. Muhammad Fajar, Fitriyanty Dwi Lestary, Jumadi Mabe Parenreng
Multi-object tracking (MOT) in crowded scenes remains challenging due to short-term occlusion, local overlap ambiguity, and detector-induced errors. This study proposes a hybrid MOT framework that combines a custom-trained YOLO26 detector, ROI-based optical flow, and monocular pseudo-depth cues to improve identity preservation under imperfect detections. To provide a more complete evaluation, the framework was tested under two scenarios: ground-truth MOT17 detections, representing ideal observations, and YOLO26-generated detections, representing practical end-toend tracking-by-detection conditions. The proposed method was compared with SORT, PD-SORT, DeepSORT, and ByteTrack using standard MOT metrics, including MOTA, IDF1, HOTA, and FPS. Under the ground-truth scenario, the proposed method achieved 98.54% MOTA, 89.51% IDF1, and 91.50% HOTA, with only 8 ID switches, indicating strong identity consistency under ideal detections. Under the YOLO26 detector scenario, the method achieved 22.59% MOTA, 15.90% IDF1, and 20.30% HOTA, with 465 ID switches. Ablation results further showed that adding optical flow and pseudo-depth improved the internal baseline from 1 9. 54% to 22.59% MOTA, 12.83% to 15.90% IDF1, and 17.21% to 20.30% HOTA, while reducing ID switches from 742 to 465. The results indicate that the proposed cues are most beneficial under detector noise, where they improve identity stability and reduce fragmentation, although with a clear runtime trade-off. These findings highlight the importance of detector-tracker coupling and position the proposed framework as a targeted robustness-oriented solution for short-term ambiguity rather than a general-purpose state-of-the-art MOT replacement. © 2026 IEEE.
Computer Engineering, State University of Makassar, Makassar, Indonesia; Electrical Engineering, State University of Makassar, Makassar, Indonesia
