Efficient Gait Recognition with SEGait-ConViT: A Vision Transformer Model Enhanced by Sparse Edge Information

Accurate identification of a person from huddled scenarios like surveillance camera feeds is one of the arduous techniques in the domain of computer vision. The gait of a person is a reliable biometric recognition technique, enabled by its capability to perceive a target from a few meters away. Gait analysis can recognize a person from 5 to 10 meters away, depending on environmental factors. Vision Transformers, an alternative approach that can overcome the shortcomings of convolutional neural networks, mandate time-intensive, pricey pre-training on massive datasets. This research proposes a Vision Transformer model for gait-based human recognition. The proposed Vision Transformer model uses gait energy images as input, combined with edge information and sparse regions within those edges. The SEGait-ConViT integrates random convolutions into the self-attention layers and leverages optimized hyperparameter values to improve performance. Gait3D and CASIA-B, two popular public datasets, were utilized to assess the model’s efficacy. The suggested methodology achieves state-of-the-art achievement, as per the empirical results which is 66.5% accuracy on Gait3D and 98.8% accuracy on CASIA-B dataset.