A Cross-Camera Pedestrian Retrieval Study

Existing pedestrian retrieval algorithms have many problems, and we explore and optimize the algorithms. Pedestrian retrieval is the process of combining pedestrian detection and pedestrian re-identification. The existing algorithms are designed with an end-to-end framework, and we train pedestrian detection and pedestrian re-identification separately to improve the accuracy of the algorithms separately, and then do the combination during testing. In addition, global features are used to achieve advanced results rather than local features, which are slower to train. And a lightweight pedestrian detector is created, which helps the model to run and compute efficiently. Our approach is evaluated on publicly available datasets, and the results on single-domain and cross-domain scenarios demonstrate the effectiveness of our approach.