Exploring Novel Optical Properties with Attention Mechanism for Gait Recognition

Abstract

One of the most hotly debated aspects of human biometry is gait recognition. It entails understanding human propulsion without any physical touch, which makes it an effective biometric technique because it is difficult to mimic. However, images of persons captured are frequently discovered with a complex diversity of clothing and ambient statistics, resulting in a low identification rate in many occasions. The research presents a unique framework for learning the projections of two-dimensional optical flowfields. Rich optical streams are also collected, which are then adjusted using a moving average approach to keep the dispersed information over optical maps. Finally, a post-training Attention method is used to remedy the incorrect prediction, hence improving training ability. The suggested technique specifically handles self-occlusion scenarios in Gait recognition with a higher recognition rate and is evaluated on benchmark datasets, notably CASIA-B and OUM-VLP, outperforming many other existing state-of-the-art methods.