Research On Gait Recognition Of Non-model Method Based On Deep Learning

Gait recognition is a new biometric technology.Compared with face recognition,fingerprint recognition and other identity recognition technologies,it has the advantages of long recognition distance and difficult to disguise.It has received extensive attention from domestic and foreign researchers in recent years.Compared with model-based gait recognition,non-model-based gait recognition is simple to construct,has low computational effort,high computational efficiency,and low requirements on target image quality.The gait recognition of non-model method is divided into two kinds of energy map method and video sequence feature method.The energy-like graph method generates an image from a sequence of images in a period.This image is usually a gait energy map.When the angle of the clothing of the moving object and the acquisition device changes,the contour of the gait image also becomes larger.Changes,these changes will have an impact on gait recognition.The video sequence feature method needs to input the gait contour map in a cycle in order,which loses the flexibility of gait recognition,and in actual life,it is usually difficult to obtain the gait contour map of the complete cycle due to the interference of external factors..In response to the above problems,in order to improve the accuracy of gait recognition in cross-view and multi-walk conditions and achieve the flexibility of gait recognition,this paper combines the excellent performance of deep learning to launch a non-model method of gait recognition research based on deep learning.The work is as follows:(1)A gait recognition method based on capsule network and feedback weight matrix is proposed in the gait energy graph method combined with deep learning.The traditional convolutional neural network will cause losses during the pooling process and cannot accurately reflect the relationship between the feature attributes inside the gait energy image.This paper uses a capsule network to represent features in a vector manner,and retains the relationship between feature attributes.The previous methods of highlighting the importance of different parts of the body are only suitable for still images,and the prominent parts are fixed.This paper uses a feedback weight matrix to update the input image.In the case of a layer of capsule network,the feedback weights are calculated based on the gait features output from the convolutional layer.Experiments are performed under the gait energy image database of CASIA-B and the gait energy image database of OU-ISIR.The experimental results show that Compared with previous methods,the method is good at extracting features under vertical and parallel walking conditions,and achieves an average recognition accuracy of 61.1% under the walking condition of the CASIA-B gait energy image database wearing a coat,indicating that the method is good at extracting wear Gait characteristics of a coat under walking conditions.(2)In the video sequence feature method combined with deep learning,a gait recognition method based on set pooling layer is proposed.In order to achieve the flexibility of gait recognition,the pooling layer is adopted to realize the disordered input of the gait contour map.The input gait contour map may be a discontinuous non-complete period gait contour map.In order to better gather the shallow and deep collection information,the method of multi-layer global pipeline is adopted.In this method,the horizontal pyramid pooling is improved to create a more discriminative space for gait recognition.First,pre-process the gait contour of the CASIA-B database to obtain a normalized image,and then conduct experiments.The experimental results show that this method can achieve the flexibility of gait recognition.When the input image is only 7 frames of gait contour map The recognition accuracy rate reached 80.1%.