V-HPM Based Gait Recognition

Compared with other biometrics, biometric based on gait features can be collected under long-distance and contactless conditions to achieve identity recognition under contactless and long-distance conditions. At present, gait recognition methods are still sensitive to illumination and background changes and are susceptible to noise in feature extraction, the gait template approach suffers from inflexibility and neglect of timing information in recognition tasks. In this paper, Mask R-CNN, a deep learning detection and segmentation model, is used to extract gait silhouettes and achieve effective and real-time segmentation of human gait silhouettes. We propose an improved GaitSet algorithm with a vertical-horizontal pyramid pooling module, and introduce a Softmax loss function for joint training to address the problem that the triplet loss function does not consider intra-class compactness. The proposed algorithm achieves the current more advanced recognition performance on the gait dataset CASIAB, and for gait recognition under jacket walking conditions, the improvement in accuracy is more obvious.