Lower limb gait recognition method based on knowledge distillation

The need for assistive devices such as lower limb exoskeletons is steadily growing, making quick and precise gait recognition essential for optimal operation of these apparatuses. The knowledge distillation method was applied to gait recognition in this work. First, we took participant lower limb gait data, which we then resampled with step frequency taken into account. Next, multi-layer perceptron (MLP) with distinct layers were modeled by educators and students. The process of knowledge distillation was used to improve the accuracy of participant activity recognition by transferring knowledge from the sophisticated teacher model to the lightweight student model. On the training set, the teacher model's accuracy average was 96.4%. Following the process of knowledge distillation, the accuracy of the student model was 95.48%, representing a 1.12% improvement compared to the non-distilled model, with only 0.53M parameters and shorter training duration. The results demonstrate that knowledge distillation can improve the recognition speed and stability of gait recognition models, enabling more accurate and reliable results within the same patient population, thus providing more effective support for medical diagnosis and rehabilitation treatment.