Dual-Level Freezing of Gait Recognition Algorithm

With the increasing proportion of elderly population, assistive treatment systems for Parkinson’s patients with the issue of freezing of gait (FoG) become essential. However, the scarcity of high-quality FoG data poses a challenge to the research on methods with high generalization ability. To mitigate the issue of limited data, this paper introduces out-of-distribution detection to compensate for the shortage of FoG data by using cross-domain data. Furthermore, we design a dual-level architecture with a Gaussian kernel-based label fusion module to improve the model’s generalization ability across different PD patients. The performance of the proposed models is tested on the FoG dataset, which was collected from 24 patients, provided by GYENNO Technology Co., Ltd and Ruijin Hospital in Shanghai, China. Additionally, a Raspberry Pi-based data collection system is built to further validate the model’s effectiveness in different environments. The proposed model has been proven to be superior based on the experimental results.