Research On Person Re-identification Method Based On Multi-feature Fusion

With the development of computer technology,more and more reachers pay attention to the driverless cars which combine traditional cars with artificial intelligence.Unmanned driving is also called automatic driving.Automatic driving faces many challenges and safety is an important issue.Since the automatic driving is unmanned,there is no human control,the driving decision is finally made through the judgment of the road condition information by the computer.Therefore,automatic driving has higher requirements on driving safety and the safety of pedestrians in the human-vehicle system is of vital importance.Pedestrian recognition,pedestrian re-recognition and pedestrian behavior prediction are important components of human-vehicle system safety and are of great significance to the development of automatic driving in automatic driving.Pedestrian is the key observation object and target of human-vehicle system safety.Pedestrian retrieval,re-identification,tracking and behavior analysis in video have become an important part and key technology of human-vehicle system safety.In this paper,relevant research work is carried out on three issues which are picture,video pedestrian re-recognition and pedestrian behavior prediction.The innovative results are as follows:(1)Aiming at the pedestrian re-recognition of pictures,a new biometric learning method named HSMLP is proposed in this paper.The purpose of human skeleton mutual learning is to solve the influence of background and local attitude change by using the new pedestrian local segmentation method combined with the global skeleton information proposed in this paper.Firstly,the attitude and skeleton of pedestrians are estimated by the bottom-up method and the intersection points of pedestrians are marked in the process.In order to solve the influence of background on pedestrian re-recognition,a new local segmentation method named node segmentation method is proposed in this paper to conduct local segmentation and local block matching for pedestrians.In addition,the global skeleton information is learned and matched by the bottom-up method which defines the joint distance from the twodimensional skeleton estimation of pedestrians.Finally,in order to improve the performance of the model,this paper uses biometric identification-based local matching and global bone matching for mutual learning.(2)For video pedestrian re-recognition,a triple pyramid model is established in this paper to learn action information for video-based re-recognition.Firstly,the action information extracted from the three-layer pyramid model and integrated into the appearance information.Secondly,3D convolutional neural network is used to process the fusion features to realize the re-recognition of human.The triple pyramid model divides the RGB image into R,G and B parts.Then integrates the action information of the three parts extracted by the triple pyramid model to obtain complete action information.In fusion Ⅰ stage,R,G,B action information into a complete information of motion in this paper.In fusion Ⅱ stage,we change information fusion to the exterior information can complement the appearance of the whole information.The traditional two-dimensional convolution is replaced by three-dimensional convolution.We improve the method of triple loss training parameters,and applies triple loss training to video pedestrian re-identification and update the network parameters in this paper.The triple loss of video not only includes the measured distance loss between video and the total measured distance loss within video,but also includes the action information loss between video and video and the appearance information loss between video and video.By comparing with the accuracy of MARS,ilids-vid and prid-2011 three video pedestrian data sets,the good performance of video person re-identification based on the RGB triple pyramid model was verified.(3)In view of pedestrian behavior prediction,this paper proposes a new pedestrian behavior prediction method--a behavior prediction method based on mesh bone division.The method mainly is divided into two parts.Firstly,we use the bones of the bottom-up approach to extract information,rules about elbow left,right,left and right knee and foot eight key points as the extraction behavior of key points.Then learning eight key points of distance measurement and angle measurement characteristics extracte behavior characteristics.The probability of the movement type of a single node in the next frame is determined by comparing the behavior characteristics of the 8 nodes before and after the frame.the pedestrian’s action in the next frame is determined by weighting the movement type of the 8 nodes in the next frame.In order to better evaluate the action of pedestrians in the next frame,we divide the pedestrians into grids and extracts the meshing features of the corresponding nodes.By comparing the meshing characteristics of the front and rear frames,the motion direction and speed of pedestrians can be judged.This paper proposes corresponding solutions and algorithm models for the problems in pedestrian re-recognition and behavior prediction based on pictures and video,and verifies the effectiveness of these algorithms through a large number of experiments.The research results of this paper will promote the development of unmanned driving and human vehicle system safety.