Non-local gait feature extraction and human identification

As a new human identification technology, gait recognition is receiving more and more attention in recent years. However, traditional gait recognition techniques are limited by the challenges of feature representation and extraction algorithms. In this paper, by utilizing the self-attention mechanism, we propose a novel gait-based human identification solution. Firstly, we utilize non-local neural networks (NLNN) to extract non-local features from a pair of randomly selected gait energy maps (GEIs). Secondly, based on the relationship between GEIs and various parts of the human body, the output of NLNN is horizontally segmented into three sections, i.e., strong-dynamic region, weak-dynamic region and micro-dynamic region, respectively. Thirdly, the segmented gait features are weighted ensembled by three two-class classifiers. Finally, two experiments are carried out with the OU-ISIR large population dataset and the CASIA dataset B to evaluate the proposed approach.