Human Gait Recognition Using Image Entropy Vector With Extreme Learning Machines

Muqing Deng,Jili Li,Jiangmin Tian,Xiaoping Lai,Jiuwen Cao,Zhiping Lin

2020-01-01

Abstract:The compromise between recognition accuracy and computational complexity is among the most difficult tasks in the area of pattern recognition. In order to achieve better performance in terms of both accuracy and speed, we present an efficient method in this paper to exploit the advantages of image entropy features and extreme learning machine (ELM) for human gait recognition. The proposed method consists of two phases: First, an image-entropy-based feature extraction strategy is introduced, in which row-based entropy vectors of binary contour are extracted for each walking sequence. Second, an ELM network is trained by supervised learning. A bagging algorithm is adopted to improve the stability and generalization performance of the ELM network. Repetitive sampling is performed on the training set, on which the ELM network is re-trained. The final recognition result is then determined by a majority vote rule on the multiple ELMs trained through the bagging algorithm. The computational burden can therefore be reduced in this work by using the simplified feature extraction process, effective image entropy features and the fast learning ability of ELM. This paper further constructs a unified multi-view training set containing gait features of each individual observed across different view angles, and discusses the extension of the proposed method for multi-view gait recognition. Extensive experiments on CASIA-A and CASIA-B gait databases demonstrate that the proposed method can achieve encouraging recognition accuracy in both single-view and multi-view situations with outstanding computational efficiency.