Abnormal gait recognition with millimetre-wave radar based on perceptual loss and convolutional temporal autoencoder

Walking is fundamental to normal human life. However, many people suffer from walking impairments due to various diseases that may severely affect their daily activities. Early detection of an abnormal gait can aid subsequent treatment and rehabilitation. This paper proposes a novel abnormal gait recognition method based on a perceptual loss convolutional temporal autoencoder (PLCTAE) network. It comprises upstream and downstream tasks, both of which utilise radar micro-Doppler spectrograms as inputs. The upstream task employs a convolutional autoencoder with the perceptual loss to encode and decode micro-Doppler spectrograms, achieving unsupervised pretraining and obtaining the initial parameters for the convolutional part of the PLCTAE. The downstream task fine-tunes the convolutional part of the PLCTAE through supervised training to extract spatial features from the micro-Doppler spectrograms and incorporates a bidirectional long short-term memory (BiLSTM) network to further extract temporal features, accomplishing the task of abnormal gait classification. The experimental results demonstrate that the proposed method achieves good classification performance on the self-established dataset which is collected by Texas Instruments' IWR6843ISK millimetre-wave radar and contains eight types of abnormal gaits. The generalisation performance is also validated on a public dataset from the University of Glasgow containing six types of human activities.