Infrared Human Recognition Based On Gait Energy Image

As a kind of biometric recognition technology,gait recognition has always been a research hotspot in computer vision and other fields,and has very important application value in unmanned driving,video surveillance,police criminal investigation and other fields.Gait recognition is the use of infrared thermal imager on human gait detection,unlike visible light imaging principle,the infrared image of gait is internal heat radiation to human body induction,and a light is not sensitive,ability to penetrate smoke fog,the advantages of the gait infrared image scene noise big,poor visual characteristic features,cause human body target in infrared image identification there are greater difficulties.At present,most gait recognition methods are proposed for gait images in visible light.This paper proposes a double convolution infrared gait recognition model based on gait energy map.First,the image sequence in the CASIA gait database(Dataset C)is segmented,and the original infrared gait images are enhanced by a combination of wavelet transform and CLAHE(Contrast Limited Adaptive Histogram Equalization)enhancement method.Secondly,after extracting the target human body,we utilize morphological methods(e.g.,dilation,corrosion)to deal with noise and lack of necessary details in infrared gait images.Further,the size and centroid of the human body image are determined by the normalization operation.Then,the gait energy map generated by the pre-processed infrared gait image is used for sample testing.We discuss the parameters of the network structure of each layer and the size and number of convolution kernels in the proposed model.Finally,the adaptive learning rate Adam algorithm is used to optimize the learning model.All experiments on the proposed infrared gait recognition method are implemented in Python using Pytorch.Compared with ordinary infrared images,simulation experiments show that the gait energy map can better characterize the infrared human gait.And the double convolutional network structure with a small memory footprint can consume less time than the single-layer model.Therefore,the proposed model significantly improves the recognition accuracy and stability.