Research On Person Re-Identification Algorithms In Video Surveillance

In recent years,for the great demand of security projects such as safe cities and smart cities,video surveillance systems have been widely deployed throughout the country.Large-scale video surveillance networks generate massive amounts of video data every day,the traditional method of manual data analysis is far from meeting the security demand.Thus,the intelligent video surveillance system has attracted a lot of attention from research institutions and police departments.Person re-identification is a key part of the intelligent video surveillance system,which aims to determine whether images/videos of pedestrians captured by different cameras are the same pedestrian.Since it can quickly and efficiently search and track specific pedestrians in a large-scale database,person re-identification has become a popular research field of computer vi-sion and multimedia analysis.Due to the complexity and changeability of the real mon-itoring environment,the traditional and reliable biological features are hard to obtained.Existing person re-identification methods are mainly based on appearance information for distinguishing pedestrians.Accurate and robust person re-identification is still a challenging task,due to occlusion,dramatic variations in camera viewpoint and illumi-nation etc.The dissertation studies person re-identification systematically and deeply,according to different feature information and application scenarios,different models are proposed from four directions to improve the performance of re-identification.The main research contents and innovations are as follows:(1)Appearance based person re-identification.Based on the deep learning tech-nique,this dissertation proposes two different deep network models.1.The multi-scale triplet convolutional neural network,which joints the process of feature extraction and metric learning,and extracts multi-scale appearance information from images for learn-ing discriminative pedestrian features.In addition,the network proposes an improved triplet loss to optimize the model.2.Joint contextual and attentional-comparative convolutional neural network,it utilizes the spatial context among body parts to learn context-aware representation,and uses attention-aware comparative network to learn attention-aware comparative representation.The two representations are complemen-tary and thus are learned jointly for extracting more effective pedestrian representa-tion.The experimental results show the effectiveness of the two models,achieving high recognition accuracy.(2)Appearance and semantic attribute based person re-identification.Appearance features are easily interfered by different factors,result in poor robustness.This disser-tation proposes a method of combining semantic attribute and visual appearance,and designs a novel contextual-attentional attribute-appearance network to learn discrim-inative and robust pedestrian representation,as well as improve the robustness of the model,by jointly learning semantic attributes and visual appearance of pedestrians.The experimental results on multiple image datasets indicate that the network can extract ef-fective pedestrian features,and achieve high performance of re-identification.(3)Appearance and temporal information based person re-identification.Com-pared to static images,video sequences contain richer temporal information.In order to make full use of such information,this dissertation proposes a dense 3D convolu-tional neural network.It combines 3D convolution operation and residual block,which is capable of automatically learning effective spatio-temporal and appearance features from video sequences.Meanwhile,it jointly uses an identification loss and a center loss for learning more discriminative pedestrian features.The experimental results on two common video datasets show the effectiveness of the proposed network.(4)Cross-domain person re-identification.The existing methods have poor gen-eralization ability,which makes it difficult to practical application.This disserta-tion proposes a novel adaptive transfer network for effective cross-domain person re-identification.It looks into the "black box" of domain gap,decomposes the complicated cross-domain style transfer task into a set of factor-wise sub-transformers,and adap-tively paints more precise style transferred images according to the influence of different factors for enhancing the generalization ability of the model.The experimental results on multiple image datasets indicate that the proposed method can effectively bridge the domain gap and improve the performance for cross-domain person re-identification.