PoseMapGait: A model-based gait recognition method with pose estimation maps and graph convolutional networks

Abstract

Gait recognition is a particularly effective way to avoid the spread of COVID-19 while people are under surveillance. Because of its advantages of non-contact and long-distance identification. One category of gait recognition methods is appearance-based, which usually extracts human silhouettes as the initial input feature and achieves high recognition rates. However, the silhouette-based feature is easily affected by the view, clothing, bag, and other external variations. Another category is based on model-based, one popular model-based feature is extracted from human skeletons. The skeleton-based feature is robust to many variations because it is less sensitive to human shape. However, the performance of skeleton-based methods suffers from recognition accuracy loss due to limited input information. In this paper, instead of relying on coordinates from skeletons, we exploit that pose estimation maps, the byproduct of pose estimation. It not only preserves richer cues of the human body compared with the skeleton-based feature, but also keeps the advantage of being less sensitive to human shape compared with the silhouette-based feature. Specifically, the evolution of pose estimation maps is decomposed as one heatmaps evolution feature (extracted by gaitMap-CNN) and one pose evolution feature (extracted by gaitPose-GCN), which denote the invariant features of whole body structure and body pose joints for gait recognition, respectively. Our method is evaluated on two large datasets, CASIA-B and the CMU Motion of Body (MoBo) dataset. The proposed method achieves the new state-of-the-art performance compared with recent advanced model-based methods.