A Lightweight Deep Learning Algorithm for Identity Recognition

The challenges in current WiFi based gait recognition models, such as the limited classification ability, high storage cost, long training time and restricted deployment on hardware platforms, motivate us to propose a lightweight gait recognition system, which is named as B-Net. By reconstructing original data into a frequency energy graph, B-Net extracts the spatial features of different carriers. Moreover, a Balloon mechanism based on the concept of channel information integration is designed to reduce the storage cost, training time and so on. The key benefit of the Balloon mechanism is to realize the compression of model scale and relieve the gradient disappearance to some extent. Experimental results show that B-Net has less parameters and training time and is with higher accuracy and better robustness, compared with the previous gait recognition models.