Research On Deep Learning Based Person Re-identication Technology

Person recognition has always been one of the hot topics in the field of computer vision artificial intelligence.In recent years,the rapid development of artificial intelligence,intelligent robots,automatic robot driving,intelligent video surveillance systems and other applications continue to make the pedestrian recognition problem more and more important.Pedestrian recognition The use of computer vision technology and artificial intelligence technology to analyze the pedestrian targets that appear under the camera to obtain their movement status,which has a great effect on crime prevention,traffic control and automatic driving.The essence of pedestrian re-recognition in a non-overlapping view is to determine whether a pedestrian under the surveillance of a camera appears under another camera and can be continuously tracked.In the real scene,the pedestrian appearance collected by the monitoring system is greatly different due to the change of the angle of the camera,the influence of the background,the change of the light intensity,the change of the pedestrian posture and the appearance of the obstruction during the pedestrian movement,Re-identification problems have brought great challenges.In the training process of Convolutional Neural Networks(CNN),this paper adds pedestrian attribute in the training process,and uses the training results to identify pedestrians.In this paper,CNN is used to detect pedestrian posture joints and a pedestrian preprocessing dataset is constructed.Firstly,the CPM(Convolutional pose machines)framework is used to detect the joint point of the pedestrian in the original data set.Then mapping the joint point information of pedestrians into a pedestrian PoseBox frame.Secondly,we train a property-based pedestrian re-identification network model using the PoseBox dataset and ResNet-50.In pedestrian matching,the pedestrian PoseBox box is firstly extracted for preprocessing,and then the model is used to extract the depth features of the pool 5 layer and perform kernel function learning.The original linear feature space is projected to a more discriminative nonlinearity feature space,after which an independent measure matrix is learned.In order to solve the over-fitting problem,the measure matrix obtained by learning is regularized,and then a similarity measure function is obtained for measure of pedestrian similarity.The proposed algor ithm can effectively extract robust pedestrian characteristics,and achieved good experimental results on different data sets.