Scholarlink

<p><b>Abstract:</b> In complex multi-object tracking (MOT) scenarios, targets not only frequently undergo occlusions but also often exhibit similar appearances. However, most existing tracking algorithms primarily focus on addressing the occlusion problem. To tackle the joint challenge of target occlusion and appearance similarity in crowded scenes, this paper proposes a Density-layered association-based multi-object tracking algorithm with Pose information (DP-SORT). The proposed algorithm first introduces a Density Estimation Method (DEM), which quantifies the level of dense occlusion for each detection box in 2D images by analyzing the manifestation of occlusion as the stacking and clustering of bounding boxes. This provides a quantitative basis for the precise sparsification of target sets. Subsequently, a Density-Layered Association (DLA) strategy is proposed, which stratifies the detection boxes in each frame based on their density values and performs data association sequentially from the lowest to the highest density. This enables progressive removal of occlusions and alleviates crowding-induced tracking difficulties. To further enhance tracking accuracy under appearance similarity conditions, a Pose-Assisted Tracking (PAT) method is incorporated. This method leverages the MoveNet pose estimation module to extract pose features of targets, thereby improving their distinguishability and localization precision. The proposed DP-SORT algorithm achieves competitive performance on public benchmark datasets including DanceTrack, MOT20, and MOT17, demonstrating its effectiveness and robustness in complex tracking scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DP-SORT:Research on a density-layered association MOT algorithm integrating human pose information

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Haifeng Sang,
  • Xiufen Fu

摘要

Abstract: In complex multi-object tracking (MOT) scenarios, targets not only frequently undergo occlusions but also often exhibit similar appearances. However, most existing tracking algorithms primarily focus on addressing the occlusion problem. To tackle the joint challenge of target occlusion and appearance similarity in crowded scenes, this paper proposes a Density-layered association-based multi-object tracking algorithm with Pose information (DP-SORT). The proposed algorithm first introduces a Density Estimation Method (DEM), which quantifies the level of dense occlusion for each detection box in 2D images by analyzing the manifestation of occlusion as the stacking and clustering of bounding boxes. This provides a quantitative basis for the precise sparsification of target sets. Subsequently, a Density-Layered Association (DLA) strategy is proposed, which stratifies the detection boxes in each frame based on their density values and performs data association sequentially from the lowest to the highest density. This enables progressive removal of occlusions and alleviates crowding-induced tracking difficulties. To further enhance tracking accuracy under appearance similarity conditions, a Pose-Assisted Tracking (PAT) method is incorporated. This method leverages the MoveNet pose estimation module to extract pose features of targets, thereby improving their distinguishability and localization precision. The proposed DP-SORT algorithm achieves competitive performance on public benchmark datasets including DanceTrack, MOT20, and MOT17, demonstrating its effectiveness and robustness in complex tracking scenarios.