TransGait: Vision Transformer Based Gait Recognition Network

Gait, the distinctive walking pattern of an individual, stands out as a human modality that can be effortlessly captured even with low-resolution cameras from a considerable distance. Gait patterns are unique to each individual and hard to fake. Gait recognition is a cutting-edge biometric technology that focuses on identifying individuals based on their distinctive walking patterns. Therefore, its fields of application span a number of domains, encompassing healthcare, robotics, automation, as well as security and access control. The origin of this investigation stems from challenges faced by high school and college residents contending with forgotten identification credentials, compounded by the intricacies of facial recognition in the context of the COVID-19 pandemic, especially in instances involving mask-obscured facial features. This paper proposes Transformer based Gait Recognition Network (Occlusion-TransGait), which combining temporal Vision Transformer (VIT) and Swin Transformer based structure and occlusion masked datasets. Utilizing the Swin Transformer model, the study achieved enhanced precision, with improvements of 0.4% for NM, 1.0% for BG, and 1.4% for CL. Furthermore, this paper presents a novel open-set detection technique for recognizing potential threats arising from unidentified individuals. Finally, this research embraces an inventive strategy that integrates human detection, tracking, and segmentation to create an end-to-end pipeline well-suited for implementation on the NVIDIA Jetson platform. This research promises a transformative impact on gait recognition systems, enhancing their effectiveness and security, and paving the way for stronger and more efficient biometric solutions.