LSA3D: Lightweight Separate Asynchronous 3D Convolutional Neural Network for Gait Recognition

Gait recognition is gaining attention as an emerging biotechnology for identifying subjects from a remote distance. Existing works have confirmed that modeling the motion trajectory of the subject is the key to gait recognition. Considering the motion continuity in the gait sequences, 3D CNN is a suitable and powerful tool for extracting spatiotemporal features. However, directly stacking 3D convolutions to extract sequence features will increase the model complexity and the number of parameters. To address the above issues, we propose a lightweight separated asynchronous 3D convolutional neural network (LSA3D) for gait recognition. Unlike common stacked 3D architecture, our network adopts the reverse pyramid design pattern, i.e., the top uses 3D convolution to extract internal features with rich spatiotemporal information, and the bottom uses separated asynchronous convolution to exploit deeper interactions between features. Experiments on the CASIA-B and OUMVLP gait recognition datasets show that LSA3D outperforms the state-of-the-art gait recognition methods.