DCNN-SVM-Based Gait Phase Recognition With Inertia, EMG, and Insole Plantar Pressure Sensing

Gait phase detection holds great importance in the field of human activity detection and medical rehabilitation, but at present, gait recognition technology still has the disadvantages of insufficient portability and high cost. Gait recognition based on pressure pad measurement limits the subjects to a small space, while gait recognition based on inertial measurement units (IMUs) or electromyography (EMG) sensing is difficult to improve recognition performance due to single-sensor or single-feature extraction model. Besides, complex neural networks also have some problems, such as lack of timeliness and redundant parameters. In this study, multisensor data and multiple classifiers are utilized to extract abundant gait features. We have developed a flexible insole utilizing fiber Bragg grating (FBG), combined with acceleration sensors and EMG sensors to accurately recognize various gait phases, including loading response (LR), mid-stance (MS), terminal stance (TS), preswing (PSw), and swing. Feature extraction from plantar pressure data and inertial data was performed using an error-correcting output coding support vector machine (ECOC-SVM) model. In addition, a deep convolutional neural network with a squeeze-excitation module (DCNN-SE) was employed for feature extraction from EMG data. Finally, gait phase detection based on multisensor data and multiple classifiers was achieved through weighted discrimination algorithms. Experimental results demonstrate that the prepared flexible insole effectively detects changes in plantar pressure with exceptional robustness and maintainability. Furthermore, the fusion model significantly enhances gait phase detection accuracy up to 96.00%, which lays a foundation for intention perception during human walking and robot walking control.