.png){kind=icon}
.png){kind=icon}
Return
Pedestrian multi-object tracking combining appearance and spatial characteristics
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126772.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Pedestrian multi-object tracking algorithms play a crucial role in enhancing public safety, optimizing traffic flow, and managing smart cities. However, existing algorithms often suffer from identity confusion in occlusion scenarios, leading to tracking failure. To address this issue, this paper introduces ASTrack, an algorithm that combines both appearance and spatial features to achieve stable tracking in occlusion scenarios. ASTrack uses deformable convolution to better extract pedestrian appearance features, enhancing its ability to distinguish pedestrians in occlusion scenarios by incorporating height information into spatial features. The algorithm also adopts an adaptive association strategy, which dynamically adjusts the balance between appearance and spatial features based on the level of occlusion, leading to improved tracking performance. The proposed algorithm was evaluated in diverse occlusion scenarios, including different lighting conditions and camera angles, attaining a tracking accuracy of 79.3 on the MOT20 dataset, with identity switches reduced by 314, down to 909 compared to the baseline. These results demonstrate that ASTrack notably enhances the robustness of pedestrian multiobject tracking algorithms in occlusion scenarios.
Keywords:
Pedestrian multi-object tracking
Occlusion scenarios
Deformable convolution
Height Iou
Adaptive association strategy
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W