Research On Human Gait Recognition Technology Based On Key Parts Segmentation

Along with the new wave of technological development,there are tremendous changes in people's social life.With the gradual improvement of social productivity,people also have high expectations for the improvement of the happiness.One of the most important aspects is the safety of personal and property.Security protection is a social supply that came into being under the stimulus of huge social demand,and its product form is constantly undergoing innovation and progress.Computer vision,as an important information acquisition method,is a process of integrating and processing visual information acquired through pictures or video images,so as to achieve the final recognition of visual information.In the past many years,the field of computer vision has carried out early explorations.Later,with the accumulation of early algorithms and advances in hardware technology,especially the continuous improvement of computing platform performance represented by GPU,the field of computer vision has made considerable progress.As an important part in the field of computer vision,gait recognition has a wide range of application scenarios due to its long-distance,non-contact,non-facial feature sensitivity and other characteristics.In the past,the accuracy of gait recognition was not high,and most of the recognition algorithms were improved on the basis of existing data sets.Based on this,the video data of CASIA-B database is selected as the original data source;by comparing and implementing different target positioning algorithms,the Faster-RCNN network model is selected and improved,and the anchor point design of the region proposal network is improved,and the target positioning is improved.As a result,the accuracy of gait recognition is improved;3 types of segmentation methods for key parts of the human body are proposed,and 5 types of gait recognition data combination methods are designed as input sources.The gait recognition network model is trained to improve the accuracy of gait recognition;The design and implementation of the gait recognition platform and the two-person gait recognition and comparison platform make the gait recognition technology more practical and close to the actual use scene.First,choose the character gait video in CASIA-B database as the original data source for corresponding model improvement and engineering practice.Through image framing,the video data is divided into image data with time series.Based on a comprehensive comparison of RCNN,Fast-RCNN,Faster-RCNN and other algorithms,the pros and cons of each party are analyzed,and Faster-RCNN is selected as the target location recognition algorithm in terms of recognition accuracy and training efficiency.And on this basis,the specific network structure of Faster-RCNN has been improved,and the recognition accuracy has been improved on the premise that the time complexity is not greatly increased.In the instance segmentation process,Mask-RCNN with better segmentation accuracy and moderate recognition efficiency is selected as the segmentation algorithm.In the network training session,the COCO database labeled by DensePose-COCO is used as the training data source for image segmentation.After that,the key parts and other parts are separated from the foreground and the background by means of binarization.Three key parts segmentation methods are proposed,and based on this,five gait recognition data combination methods are produced.Analyzed a series of early gait recognition algorithms based on manual feature extraction and recent gait energy maps based on deep convolutional networks,human skeleton key points and other related algorithms,combined with GaitSet's image collection idea and attention mechanism,by using the previous step The segmented image sequence diagram with key part features is injected into the convolutional neural network in the form of a collection,and the obtained features are then subjected to feature extraction,collection pooling,horizontal pyramid mapping,and feature comparison,and finally the gait is obtained Recognition results.The final result shows that the recognition accuracy has been improved to a certain extent.By integrating the aforementioned related algorithms,two sets of practical gait recognition tools have been developed,including a single-person gait recognition platform and a two-person gait recognition and comparison platform.The two-person gait recognition and comparison function has a wide range of practical application scenarios.,Can help police or other statutory public servants to determine the target.In the case of the limited gait database,it can industrially and batchly determine whether the target in the two videos is the same person,which can greatly improve the police The efficiency of other civil servants.