Deep mutual learning network for gait recognition

Human identification plays a significant role in ensuring social security. However, face-based and appearance-based retrieval methods are not effective in monitoring due to the long distance and low camera resolution. Compared with other biological characteristics, the gait of humans has a strong discriminating ability even at long distance and low resolution. In this paper, the deep mutual learning strategy is applied to gait recognition, and by training collaboratively with other networks, the generalization ability of the network is improved simply and effectively. We use a set of independent frames of gait as input to two convolutional neural networks. This method is unaffected by frame alignment and can naturally integrate video frames of different walking conditions (e.g. different viewing angles, different clothing/carrying conditions). At the same time, the set can extract gait features from incomplete gait cycles due to occlusion. A mutual learning strategy can improve the running speed appropriately and realize the compactness and accuracy of the model. Two convolutional networks learn simultaneously and solve problems together. To evaluate the method's performance, we compare it to several methods on the CASIA and OU-ISIR gait databases, and construct different sets of gaits with incomplete periods to compare the accuracy of identification with them and the complete gait set. Experimental results show that the method is effective.