Deep Learning Based Multi-Target Multi-Camera Tracking System
Chenwei Dong, Junlei Zhou, Weipeng Wen, Si Chen · 2022
Multi-target multi-camera tracking has lately received great attention for computer vision. However, the traditional methods are difficult to be applied to the industry due to their low efficiency and high cost. In order to overcome the above problem, we design and implement a novel deep learning based multi-target multi-camera tracking system, termed MTMCT, which combines three sub-modules, i.e., person detection, single-camera tracking (SCT), and person re-identification (Re-ID) to achieve multi-camera pedestrian tracking and cross-camera pedestrian retrieval. In this system, we adopt the tracking-by-detection strategy and introduce the coordinate attention mechanism in the person detection sub-module, which not only considers the relationship between the channels but also pays attention to the position information in the feature space. In addition, we employ a fast inner product search based on Möbius transformation to improve the efficiency of this system. Experiments demonstrate the effectiveness of our system in real applications.