@

초록

Object perception using multi-view cameras is essential for intelligent systems operating in indoor environments such as warehouses, retail stores, and hospitals. Traditional multi-target multi-camera (MTMC) detection and tracking approaches typically depend on 2D object detection and single-view multi-object tracking (MOT), without properly exploiting 3D spatial information. While point cloud-based 3D detection models offer high accuracy, their adoption is hindered by expensive deployment requirements. Besides, most existing MTMC tracking methods are designed for daytime scenarios, neglecting the challenge posed by lowlight conditions. In this paper, we introduce DepthTrack, a novel MTMC tracking framework for all-day tracking without requiring high-precision 3D sensors to produce 3D bounding boxes. At the core of DepthTrack is our Tracklet-Cluster Mapping (TCM) strategy, which seamlessly integrates 3D object clusters with bird's-eye view (BEV) tracklets to achieve robust 3D tracking. Experiments conducted on the AICity'25 dataset validate the strong generalization of DepthTrack across diverse lighting conditions. In the 2025 AI City Challenge Track 1, our team secured the second position with an accuracy (HOTA) of 63.1396. The code will be released at https://github.com/SKKUAutoLab/AIC25_Track_01

키워드