Human gait recognition using STFT-CNN approach based on wearable biomechanical sensor data

Human gait pattern recognition has been an increasingly popular application of biomechanical sensors in recent years. Wearable inertial measurement unit (IMU) and goniometer (GON) sensors are crucial for precise human activity recognition, with their placement and number significantly affecting assessment and clinical applicability. This study presents a novel multi-channel sensor signal processing method that utilizes short-term Fourier transform (STFT) and convolutional neural network (CNN)-based deep learning (DL) approach (STFT-CNN) for automatic recognition of human walking speed (WS) using wearable biomechanical sensor signals. In this approach, the gait STFT images are applied to the DL network input and trained with a 2D-CNN model for the classification of WS. This data from 22 healthy individuals is analyzed using an 80:20 train test split approach, and the model reliability is evaluated. The single-input “IMU-5s” and “GON-NoSeg” CNN models achieved 0.910 and 0.814 accuracy, respectively, while the multi-input “Multi-5s” and “Multi-NoSeg” CNN models, incorporating GON sensor data, resulted in 0.842 and 0.828 accuracy, respectively. The findings are presented in a comparative manner with those of preceding studies. The proposed approach has significant potential in assisting physicians in the diagnosis, progression, and assessment of gait disorders. Future studies should include gait characteristics such as EMG and kinematic analysis to enhance the generalizability and clinical utility of the model in patient group data containing various gait patterns.