Enhancing multi-object tracking efficiency through result-guided feature extraction and query filtering

To address the issues of slow object detection and inefficient feature extraction in existing multi-object tracking tasks, we propose a novel result-guided multi-object tracking algorithm. By leveraging historical tracking and detection results, we design a trace-back query filtering module to dynamically adjust the appearance feature extraction strategy, significantly reducing computational overhead and improving tracking efficiency. Combined with a shared structure learning-based detection-feature extraction module, our approach not only enhances feature interpretability but also improves tracking robustness and accuracy. Experimental results on the DIVO dataset demonstrate that compared to the baseline model, our algorithm achieves notable improvements in key metrics, including HOTA (+ 3.64%), MOTA (+ 4.45%), and IDF1 (+ 5.32%), while also enhancing real-time tracking performance. Our code is publicly available at: http://github.com/Yushen-W/ReMOT.