Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (21): 213-222.DOI: 10.3778/j.issn.1002-8331.2110-0231

• Graphics and Image Processing • Previous Articles     Next Articles

Pedestrian Multi-Object Tracking Algorithm with Strengthened Re-Identification

WANG Liming, SUN Jun, CHEN Qidong   

  1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2022-11-01 Published:2022-11-01



  1. 江南大学 人工智能与计算机学院,江苏 无锡 214122

Abstract: In multi-object tracking tasks, identity re-identification(Re-ID) generally depends on the quality of detection performance. Detection deviations cause the Re-ID feature to blur, thereby reducing recognition accuracy. Especially in complex scenes such as scale transformation and frequent occlusion, Re-ID is not robust, and the effect of multi-object tracking is poor. To strengthen the Re-ID performance, a pedestrian multi-object tracking algorithm is proposed. This algorithm uses CenterNet to detect objects by predicting center point heatmaps, and designs a detection deviation loss to further constraint response value on the predicted heatmaps, which alleviates the blurry problem of Re-ID features bring by inaccurate detection. Moreover, to strengthen Re-ID robustness, the algorithm implements a novel strategy to adaptively increase the learnable Re-ID features around objects center. As a result, it improves the quality of Re-ID features, and makes Re-ID task less dependent on the detection accuracy. Experimental results on MOT16 and MOT17 benchmarks show that the proposed algorithm can effectively improve the Re-ID performance in comparison with other SOTA methods, which achieves real-time tracking with 25.6 FPS.

Key words: multi-object tracking, re-identification(Re-ID), center point detection, real-time

摘要: 在多目标跟踪任务中,重识别(re-identification,Re-ID)效果通常依赖于检测性能的好坏,检测偏差会导致Re-ID特征模糊,从而降低重识别精度。特别是在尺度变化和频繁遮挡等复杂场景下,Re-ID鲁棒性不高,多目标跟踪效果较差。针对该问题,提出一种加强重识别的行人多目标跟踪算法。该算法以CenterNet为检测器,通过预测目标中心点热力图来检测目标位置,并设计检测偏差损失加强对预测热力图响应值的约束,以缓解因检测不准确导致的Re-ID特征模糊问题。为提高Re-ID鲁棒性,提出Re-ID可学习特征动态扩充策略。该策略通过自适应扩充目标中心的Re-ID可学习特征来提高特征质量,并减小Re-ID对中心点检测精度的依赖。在MOT16和MOT17测试集上进行验证,结果表明,算法能有效提升Re-ID性能,与主流算法相比具有更好的跟踪效果,且兼顾了实时性,达到25.6 FPS。

关键词: 多目标跟踪, 重识别, 中心点检测, 实时