FSGait: Fine Grained Self-Supervised Gait Abnormality Detection

Bingzhi Duan, Xiaoyue Wan, Xu Zhao; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 2248-2264

Gait Abnormality Detection (GAD) plays an important role in diagnosing diseases associated with abnormal gait patterns. However, existing works are limited in generalization capability and granularity, which indicates they detect only the types of abnormalities seen during training and sequence-level gait anomalies. The reason is the restricted variety of abnormalities in datasets and the reliance on supervised learning algorithms. Therefore, we propose a Fine-grained Self-supervised Gait Abnormality Detection method (FSGait). We divide gait abnormality into two sub-problems: postural anomaly and temporal anomaly, which are solved by two designed modules, Gait Reconstruction Module (GRM) and Gait Prediction Module (GPM). The two modules are trained in self-supervised way on normal gait data. In this way, they capture normal gait patterns to distinguish abnormalities, thereby enhancing the generalization capability. For fine-grained detection, three-level (Sequence, Frame and Joint) abnormal detections are achieved with the intermediate results of these two modules. FSGait has a high degree of granularity and holds significant potential for aiding medical diagnosis and automating disease detection. Experiments on two datasets show that FSGait achieves state-of-the-art performance in frame-level GAD, while maintaining high sequence-level GAD accuracy. The joint-level detection results are presented with visualization.