A Study of Gait Recognition Algorithm based on Multi-Level Pooling Residual Network

In this paper, we propose a gait recognition algorithm based on multi-level pooling residual network to address the problem that traditional gait recognition algorithms have reduced recognition ability and low cross-view accuracy due to variable factors such as clothing transformation and carrying object transformation. The algorithm uses parameters such as optimal number and size of feature maps and low error rate of each layer after training optimization as the input of the network, and combines Multi-level pooling (MP) to extract effective gait recognition features by fixing the output dimension and reducing the impact of various disturbances on recognition accuracy. The gait images from CASIA-B dataset and homemade dataset are used as research objects to experiment the gait energy map features from different viewpoints and different clothing or handheld objects. The results show that the improved network can achieve a correct rate of 86%, which is feasible.