Multi-pedestrian Tracking and Person Re-identification

Lightweight Multimodal Domain Generic Person Reidentification Metric for Person-Following Robots

“…In another work in [ 18 ], tracking with person reidentification is learned, where the learned tracker tracks the target person in traffic scenes. Similarly, in [ 19 ], the authors follow the tracking-by-detection method, where the CNN-based feature matching is used to identify the target person in consecutive frames, and thus, the tracking of the target is done.…”

“…Similarly, in [ 19 ], the authors follow the tracking-by-detection method, where the CNN-based feature matching is used to identify the target person in consecutive frames, and thus, the tracking of the target is done. The frameworks in [ 18 , 19 ] are simple, but these works require fine-tuning the reidentification metric every time for for every unseen domain, and thus have low generalization. Other works including [ 20 , 21 ] solve tracking in the multicamera network for smart city applications.…”

“…The authors believed for the large-scale scenario that unsupervised reidentification requires no label of identities for training, and thus perform better than supervised metrics. Although the above works [ 17 , 18 , 19 , 20 , 21 , 22 ] used the learned reidentification metric for visual tracking, the methods are not intended for tracking the target on mobile robots. In order to track the target on mobile robots, lightweight and efficient reidentifications are needed; therefore, in the next subsection, we cover the recent works that specifically learn the reidentification metrics for tracking the target in real time and on mobile robots.…”