DisGait: A Prior Work of Gait Recognition Concerning Disguised Appearance and Pose

Shouwang Huang, Ruiqi Fan, Shichao Wu

Abstract: Gait recognition, capable of identifying humans at a distance without the cooperation of the subjects, has significant security applications. As gait recognition edges closer to practical application, addressing various challenges posed by different scenarios becomes increasingly essential. This study focuses on the issue of disguises. Specifically, it introduces a novel benchmark gait dataset (DisGait) for investigating the performance of gait recognition methods under disguised appearance and pose, which has not been explored in contemporary gait-related research. The dataset consists of 3,200 sequences from 40 subjects under various walking conditions, including normal clothing (NM), thick coats (CO), different types of shoes (SH), and uniform white lab gown (GO). The primary distinction between this dataset and others lies in varying degrees of appearance and pose disguises. We performed gait recognition with various state-of-the-art (SOTA) silhouette-based and skeleton-based approaches on the DisGait dataset. Experimental results show that all of them are unable to achieve a satisfactory score with the average accuracies falling below 40%, which substantiated that disguised appearance and pose are important for gait recognition. Moreover, we evaluated the influence of sequence frames and extensive occlusions on gait recognition, analyzed the attributes and limitations of current SOTA approaches, and discussed potential avenues for improvement.