Pose-based Gait Cycle Detection

Abstract

For video-based pedestrian re-identification, spatial and temporal alignment is very important. It is helpful to find the most discriminative part of the video. Spatial alignment is commonly used to address these issues by treating the appearance of different body parts independently. In this paper, we pay attention to the temporal alignment problem. In the previous approaches, temporal alignment usually achieved in a gait cycle. That means the gait cycle detection is the first step. We proposed a posed-based method to detect the gait cycle, which could obtain prominent and accurate gait cycle. Particularly, giving a video sequence we take advantage of the latest pose estimation network to get the human skeleton, as well as the position of the key points. Then the distance of two ankles is calculated. By the variation of distance with the sequence, we could get the accurate gait cycle. It is helpful for the video-based pedestrian re-identification.