.png){kind=icon}
.png){kind=icon}
Return
A novel pipeline for tunnel multi-object tracking integrating cross-modality and motion model
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127640.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Tunnel multi-object tracking (TMOT) is oriented towards sensing and tracking objects in tunnel scenes. Vanilla methods use a single optical device to accomplish this task. However, when object occlusion and clustering occur, the spatial and appearance information provided by the optical device will be simultaneously weakened due to the high degree of overlap between objects. In this research paper, we introduce a cross-modal pipeline that can efficiently address the long-standing challenges in TMOT. First, the information interaction between RGB camera and Radar video is efficiently managed by introducing cross-modality coordination transformer (CCT) and temporal interactions transformer (TIT). The accurate bimodal fusion lays the foundation for our subsequent process. Next, we employ the cross-scale fusion module (CSF) to enhance multi-scale perception and refine the performance of global structure mining. When tracking fails due to drift or object invisibility,tracklet candidate proposals across the entire image can provide track reset opportunities. Specifically, we propose radar-guided confidence modeling (RGC), a decoupling and guidance mechanism that compensates for weakened spatial and appearance cues. In addition, to avoid ambiguity in metrics such as intersection over union (IoU), we incorporate height state and re-identification (ReID) techniques to strengthen associations. Our approach is simple, online, and real-time, and demonstrates impressive performance in a large number of benchmarks.
Keywords:
Multi-object tracking
Intelligent transportation systems
Tunnel
Radar guide
Radar-vision fusion
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available