Abnormal Gait Recognition in Real-Time using Recurrent Neural Networks

Abnormal gait recognition plays an important role in diagnosis of musculoskeletal disorders. Suspicious walking behaviours should be detected as early as possible, and possibly in real-time, in order to prevent further deterioration of any part of the body. Analysis tools should provide useful and accurate information via convenient setup procedures. In this work, we conduct a system for recognizing gait abnormalities in real-time where an input is an image frame captured by a single RGB camera at any instance. We view abnormal gait recognition as a time-series problem which requires learning long-term dependencies. Hence, the system is presented with variants of Recurrent Neural Networks (RNNs). The proposed deep neural network model involves extracting 135 human body key points using OpenPose prior to performing recognition task which are quantitatively evaluated based on a simple RNN, Long Short-Term Memory Network (LSTM), and Gated Recurrent Unit Network (GRU). Each deep neural network has a model accuracy about 73.4%, 82.8%, and 81.6%, respectively. According to the confusion matrices of different predictive models, LSTM and GRU provide less confusing predictive results than that of a simple RNN. Therefore, we have discovered that deep neural network based on LSTM is, by far, the suitable model to recognize abnormal gaits due to the high model accuracy with less training and inference time.