Gait Recognition from Fisheye Images

摘要

Gait recognition has been a hot topic of extensive research in video-based surveillance and forensics. Compared with traditional rectilinear cameras mainly used in existing studies, fisheye cameras have a wider field of view, and hence are more suitable for gait recognition applications in navigation robots, which enables more flexible and free surveillance scenarios. In this paper, to the best of our knowledge, we propose the first framework for gait recognition from images captured by fisheye cameras. To deal with severe image distortion and partial body occlusions induced by fisheye cameras set at lower heights, we combine a set of preprocessing procedures with a state-of-the-art model-based gait recognition method. Specifically, an input fisheye image is first expanded into a panoramic view before pedestrian detection. A person-dependent gnomonic projection is then applied to the detected human region for distortion correction. Next, background regions are attenuated to improve human model fitting accuracy in complex outdoor scenes. The resulting rectified image sequence is finally fed into the gait recognition network for human model estimation and gait feature extraction. To validate the performance, we collect a real fisheye image gait dataset with various views and capture scenarios, including simplified indoor and challenging outdoor scenes. Various experiments on the collected dataset demonstrate the effectiveness of the proposed method.