Publication

Abstract : Applications for gait recognition are numerous especially in security surveillance. However, due to the variety of individual walking behaviours and the complexities of external variables during data gathering, gait identification continues to face several obstacles. Among these, shallow learning-based gait recognition algorithms struggle to attain the correct rate of recognition crucial by numerous applications, while the volume of gait training data available cannot match the demands of deep learning-based model training. In order to comply with the problem outlined above, this work offers a visual gait detection system based on an attention-based multi-scale convolutional network with sequential learning. As a first step, the approach takes multi-recurrent gait energy images (MR-GEIs) as an input in a frame by frame manner and runs each frame through the convolutional network to extract entire gait features. In the second step, the key attributes of the extracted features from multi-scale convolutions are highlighted by the attention block in order to improve prediction performance. Thirdly, the bidirectional gated recurrent unit (Bi-GRU) layer is applied to obtain the temporal relationships among the different frames of MR-GEIs in the sequential learning block. The proposed network achieves average accuracies of 93.4% on CASIA-B and 97.4% on OULP gait datasets, demonstrating superior recognition performance and improved generalizability compared to previous state-of-the-art techniques.