Enhanced hybrid prototype for few-shot class-incremental gait recognition in multi-activity scenarios using wearable sensors

Wearable devices for gait information sensing provide a reliable and robust solution for identity recognition. However, in real-world applications, gait recognition systems based on these sensing devices should adapt to diverse walking activities, tackle the challenge of limited individual data, and continuously update to recognize both old and new users. In this study, we propose a framework based on hybrid prototype enhancement to address the challenge of few-shot class-incremental gait recognition in multi-activity scenarios (FC-GRMA). Firstly, hybrid prototypes are generated by introducing auxiliary activity labels, which are more generalizable than ordinary prototypes; secondly, the prototypes are adjusted by a selective prototype enhancement module, which improves the representative and discriminative abilities of the prototypes. Finally, validation on the public dataset USC-HAD and the self-built dataset CDUT-AG shows that our proposed framework performs best in solving the FC-GRMA problem. In particular, we also discuss the effect of different numbers of activities on the model performance, and the results show that our framework effectively addresses the issue of catastrophic forgetting in multi-activity scenarios. The source code is available at https://github.com/lc321/fc-grma.git.