Research On Pedestrian Re-identification In Video Surveillance Based On Deep Learning

In recent years,with the accelerated pace of construction of safe cities,the number of surveillance cameras has exploded.In the face of massive video data,traditional manual processing is both clumsy and inefficient.Therefore,in order to solve the problem of pedestrian recognition,tracking and retrieval across cameras,pedestrian reidentification technology has been proposed.Pedestrian re-identification can retrieve and identify designated target pedestrians in a large video surveillance network,and achieve target pedestrian tracking and motion analysis by establishing identity association information of target pedestrians under different cameras.However,in the real application scenario,due to the difference of illumination,camera angle and pedestrian posture,the same pedestrian often presents different states under different cameras,which brings great challenges to the research and application of pedestrian reidentification.Based on the analysis and research of the work of predecessors at home and abroad,this paper proposes two solutions based on deep learning aim to solve the influence of illumination,camera angle and pedestrian posture changes on pedestrian reidentification.The work is compared to verify the effectiveness of the proposed method.The details are as follows:(1)A pedestrian re-identification based on feature fusion and deep ensemble network is proposed.From the aspect of feature extraction,FF-DEN not only considers the robustness of CNN global features,but also considers the local saliency of LOMO features.From the perspective of network model structure,FF-DEN fully considers the similarity between images(for the learning of metric functions),and makes full use of the category information of images,and uses the cross entropy loss function to establish the intrinsic relationship between all images.Therefore,the robustness characteristic with discrimination can be extracted,thereby improving the accuracy of pedestrian reidentification.(2)A visual attention network based on multi-layer semantic features is proposed for pedestrian re-identificationThe VAN-MSF network takes three(Anchor/Postive/Negative)pedestrian images as input.Firstly,the Feature Fusion Network is used to extract the multi-layer semantic features of the pedestrian image,and then use the Visual Attention Network to learn the pedestrian image based on the multi-layer semantic features.Finally,the explicit local features and the multi-layer semantic features are effectively combined as the final feature representation of the pedestrian image,and the network parameters are optimized by the Triplet Loss+CrossEntropy Loss.(3)Constructing an end-to-end pedestrian re-identification systemBased on the work of the third and fourth chapters,an end-to-end pedestrian reidentification network model is proposed.The network model integrates pedestrian detection with pedestrian re-identification network and cascades the two networks.At the same time,the purpose of training and optimization is to realize the automatic detection and recognition of pedestrians under the video surveillance network.Based on this,an end-to-end pedestrian re-identification system based on C/S mode is constructed.Based on the Windows environment,the MyEclipse tool and Java language are used to develop and implement the related functions of the system.In this paper,the performance of the FF-DEN and VAN-MSF network models is evaluated experimentally and compared with the current popular algorithms.The experimental results show that the two network models have higher fitting speed and higher rank1 and mAP precision,which proves the effectiveness of the proposed algorithm.