Daily Activity Recognition with Gait Segmentation and Hybrid Neural Networks

Recent research has revealed that plantar pressure can reflect human characteristics, showing great potential in the field of biometrics. This study aims to explore the feasibility and methodology of using plantar pressure for daily activity recognition. We collected more than 10,000 pressure data points from 10 volunteers using insoles equipped with 8 pressure sensors to construct a dataset. Building upon a single Convolutional Neural Network (CNN), we introduced a hybrid network model combining Long Short-Term Memory (LSTM) and Attention mechanisms, and evaluated its performance. Additionally, we analyzed the relationship between gait segmentation, the length of multiple gait cycles, and recognition accuracy. The experimental results showed that without segmentation based on gait cycles, the accuracy of classification using a single CNN was 95.0%, while the accuracy using the hybrid neural network was 97.0%. After segmenting based on gait cycles, the accuracy of classification using a single CNN reached 97.4%, and the hybrid neural network achieved an accuracy of 99.6%. These results indicate that classifying segmented gait data using a hybrid neural network can achieve superior outcomes, demonstrating significant practical value.