Pedestrian age recognition method based on gait deep learning

Abstract

A pedestrian age recognition method based on gait information and dynamic weight convolutional networks is proposed to solve the problem that face age recognition needs a close positive perspective and frontal angle which limits its application scenarios. Firstly, the method takes gait energy image as input and the gait energy image is divided into head, upper body and lower body. Then, the global and local features of gait are extracted by global local convolutional networks. Secondly, the multi-layer residual expansion networks is constructed to optimize the gait energy image, and the average pooling is used to replace the fully convolutional networks for age classification. In order to solve the problem of sample imbalance in different age groups, the dynamic weight cross-entropy loss function is utilized as the network loss function to constrain the parameter updating, so as to avoid the recognition results inclining to the class with more samples. Experimental results on OULP age dataset demonstrate the effectiveness of the proposed approach.