| Pedestrian recognition and retrieval under surveillance or video is more and more important. Due to the long distance between camera and pedestrian, normal biometric recognition methods such as face recognition, iris recognition etc. are not practical. But gait, also as a biometric recognition method, which is not easily hidden, difficult to camouflage and could get in a long distance, could be very effectively used in pedestrian recognition and retrieval within surveillance and video. A large number of studies have already shown that gait can be used as a distinguishing feature for recognition. But in the scene of video or surveillance, pedestrian would walk uncertainly, resulting in uncertainty of angle between the pedestrian and the camera, impacting the accuracy of pedestrian recognition and retrieval.On the basis of summarizing and analyzing previous studies, this paper proposes a framework based on auto-encoder for gait recognition. Based on this framework, we could transform the angle between the pedestrian and the camera to a unified view in favor of retrieving and identifying without knowing the exact angle in advance, so we could greatly improve gait recognition rate under multi-view situation. The framework uses gait energy image, which is a commonly used gait recognition feature, as input and extracts better feature to express the pedestrian gait. At the same time, the framework also could to some extent improve the recognition rate under the circumstances of carrying objects and dressing changes without knowing the pedestrian is normal walking or walking in a coat or walking with a bag in advance.This article will test on a public gait database(CASIA Gait Database B), which is collected by the Chinese Academy of Sciences Institute of Automation. Experiment results show that the framework can effectively solve the three common challenges in gait recognition, multi-view, carrying objects and dressing changes, and achieved good results. |
PDF Full Text Request