Design And Implementation Of Gait Recognition System Based On Deep Learning

As a research hotspot of computer vision,gait recognition has attracted more and more researchers' attention.Gait recognition is the use of biological features for remote identification,and has a wide range of applications in many areas,such as intelligent monitoring,motion analysis,biometric authentication,criminal investigation,virtual reality and so on.The existing human gait recognition algorithms can be divided into two categories: one is a model-based gait recognition algorithm,and the other is a non-model based.However,some existing methods still have many shortcomings in the actual operation process.For example,the accuracy of the algorithm has much room for improvement,and some computational complexity is large.This paper mainly studies the design and implementation of gait recognition system based on deep learning.It studies the key technologies of each stage of gait recognition deeply.By improving some technologies,it has achieved a fast and effective pedestrian gait recognition system.Firstly,for the features of the video surveillance,such as large data volume,high dimension and high computational complexity,this paper realizes fast and effective segmentation and execution of human gait contour by using SiamMask real-time object tracking and semi-supervised video object segmentation algorithm.Algorithm is extended to support three data input methods: video,monitor and image.It increases the flexibility of the system.However,there is a large amount of noise and data redundancy for the extracted gait contour map.In order to reduce the noise and redundant data on the gait characteristics,the obtained human gait contour map is processed correspondingly by morphological operation.Then,the image is normalized,and the standardized human body gait contour map with the same human body contour and horizontal centering is generated in proportion to prepare for the next step of recognition.In addition,by using the gait recognition method based on convolutional neural network,the set feature is extracted from the frame-level features of gait images by SP(Set Pooling)algorithm,which avoids the timing problem when data is input.The HPP(Horizontal Pyramid Pooling)method is improved,and the HPM(Horizontal Pyramid Mapping)feature mapping method is proposed to effectively extract theglobal features and local features of the image.The algorithm has the characteristics of high accuracy and fast speed.Then an improved algorithm for feature extraction is proposed for the algorithm.The nonlinear feature expansion of the set features obtained by SP improves the gait recognition extraction effect under non-standard state.The improved method is called SPBS(Set Pooling Based on Sigmoid).And the common activation function of the neural network is analyzed and compared,the parameters of the LeakyReLU activation function are fine-tuned and the PReLU activation function is used for comparison.The improved gait feature extraction method proposed in this paper is verified on the CASIA-B gait recognition database.It is proved by experiments that the improved gait recognition algorithm can effectively improve the accuracy of gait recognition to a certain extent.Finally,based on the research and optimization of each stage algorithm,an efficient and fast gait recognition system based on deep learning is designed.Under the synergy of object tracking algorithm and recognition algorithm,the system has realized gait recognition in three ways of video,image and monitoring,so as to achieve the effect of identification of pedestrians.At the same time,the system has realized real-time object detection and tracking.