| Since the national 13 th five year plan,artificial intelligence can give the security system a clearer choice of tasks and more perfect results.At present,it has gradually entered a new fast lane.And from the passive defense of the traditional security system at the time of the incident to the direction of independent prevention before the incident and independent alarm at the time of the incident.With the development of economy,the country has richer financial resources to lay out video surveillance cameras with wider range and higher definition.These high-definition cameras have made outstanding contributions to social security and public security construction,urban safety construction,informatization of relevant departments,health and regional protection and traceability.Especially during the epidemic period,the tracing work of close contact personnel assisted by the security system has achieved excellent results.Because the camera often can’t capture people’s face correctly and clearly,in some real application scenarios,relying only on face information can’t meet the actual work needs.We also need to use other information with excellent feature attribute tags,such as human posture information and clothing information,so as to promote the evolution of artificial intelligence and promote the breakthrough from simple face recognition to the stage of recognizing a "human" individual.For the pedestrian re recognition technology itself,the pedestrian re recognition task is to retrieve a specific image from a large image database.It is an image retrieval only for pedestrian images.In the process of applying pedestrian re recognition technology to practice,it also faces many difficulties.Limited by the location of a single camera and the field of view of the camera itself,a single camera cannot cover a monitoring area in an all-round way.Even if multiple cameras shoot in the same area,there will still be problems such as the difficulty of seamless connection of captured videos,which makes the existing pedestrian re recognition technology still difficult to meet the needs of real-time target matching in large-scale intelligent monitoring system.The pedestrian re recognition task will not only be affected by hardware factors,but also the practical application effect of pedestrian targets will be affected by many factors,such as clothing,viewing angle change,foreign object occlusion and so on.Although a large amount of video data has been accumulated under the acquisition of huge cameras across the country,most of these data contain little effective information about each pedestrian,and such an order of magnitude of video data is almost of no help to the pedestrian re identification task without artificial labeling.Large amount of data and insufficient effective samples,which is a typical problem of large data and small samples.This thesis focuses on identifying and solving many complex problems in the downlink scene.From the perspective of data expansion,pedestrian sample tracking model and pedestrian sample expansion method are explored.From the perspective of model improvement,an adaptive iterative method is proposed.From the perspective of system optimization,an Optimization method of Person Re-identification System Based on Time-Space Information is proposed.The main innovations include:(1)In order to solve the problem of occlusion of pedestrian samples in the task of pedestrian tracking,this thesis proposes a stable twin network tracking method STWS(Shape Robust Siamese Network Tracking Based on Semi-Supervised Learning)based on semi supervised learning,which aims to improve the positioning of targets in the image.Most of the existing semi supervised methods only locate the regions with the most significant features in the object,rather than all the relevant regions of the target,which leads to the poor performance of the model.The core idea of the STWS method proposed in this thesis is to randomly hide some areas of the target sample in the training data,so as to generate the occluded pedestrian sample based on the original pedestrian sample.The purpose is to evenly disperse the focus of the network in different local areas.When the area with the most significant characteristics is hidden,the network is forced to find other relevant parts.The STWS model proposed in this thesis can effectively spread the focus of network training from a few key positions to the whole target image,so as to better solve the problem of image occlusion.When STWS generates new occluded pedestrian samples,the original data set is greatly expanded,so as to achieve the purpose of expanding the data set.(2)According to the problem of insufficient inter frame information mining of existing pedestrian re recognition video frames,this thesis proposes a progressive pedestrian re recognition network update method RLPU(Reinforcement Learning Progressive Update Network for Person Reid)based on reinforcement learning.The characteristics of efficient use of video inter frame information through pedestrian tracking algorithm and the mapping of reinforcement learning from environmental state to action.The pedestrian track segment is obtained by the pedestrian tracking algorithm,and the pseudo labels are assigned to the pedestrians based on the track segment information.During the iterative update of the pedestrian re recognition network by reinforcement learning,the pros and cons of the pseudo tag segments are judged according to whether the operation results of the pedestrian re recognition model are significantly improved after each iteration,and the optimal pseudo tag allocation strategy is obtained,so as to dynamically screen the pseudo tags.RLPU algorithm improves the performance of existing benchmarks and has high efficiency.(3)Aiming at the problem of high labeling cost of large-scale data sets,this thesis proposes semi supervised adaptive stepwise learning SSAS(Semi-Supervised Adaptive Stepwise Learning)based on STWS algorithm and RLPU algorithm.Based on the pedestrian track segments and some pseudo labels,SSAS algorithm proposes a more global pseudo label update idea.Different from RLPU’s method of screening false labels from the iterative results of pedestrian re recognition model by reinforcement learning,SSAS uses the core idea of Kullback Leibler divergence to measure the advantages and disadvantages of false labels from the perspective of feature distribution.Gradually increase the complexity of pseudo labels,starting with simple samples and gradually increasing difficult samples.At the same time,delete the samples with weak effect and update them.The SSAS method proposed in this thesis steadily improves the recognition accuracy of pedestrian re recognition task through dynamic pseudo label screening strategy,and the superiority of this method is proved by experiments.(4)Aiming at the complex and changeable problem of pedestrian background,this paper makes further exploration based on the above methods,and puts forward the Optimization method of Person Re-identification System Based on Time-Space Information OTSI.The pedestrian tracking algorithm is used to mine the time information in the video,obtain the pedestrian trajectory segment,and expand the pseudo label video segment data set.At the same time,when the pedestrian tracking algorithm loses the target pedestrian due to external interference,the pedestrian re recognition algorithm is used to mine the spatial information in the video to obtain the pedestrian position segment,so as to improve the accuracy of the pedestrian tracking algorithm.The two complement each other,obtain the dynamic dictionary of pedestrian image data,and manage the update process by reinforcement learning to obtain the optimal update management strategy of pedestrian tracking algorithm and pedestrian re recognition algorithm.Relying on this system optimization method of time-space information combination,it can well meet the application requirements in complex scenes. |
PDF Full Text Request