Person Re-identification Across Cameras Based On Multi-features Fusion

Person Re-identification(Re-ID)is a research focus of computer vision in recent years,which has high application value in such fields as criminal investigation work of intelligent security and image retrieval.Given a person image of the surveillance video,all the images of the target are recognized across the cameras.Considering the difference of illumination,angle and pose,along with that the external pedestrians are easily influenced by the detection accuracy and occlusion and other factors,how to improve the re-identification rate of pedestrian images still faces significant challenges.According to the characteristics of pedestrian images,the design scheme of person re-identification system is proposed in this thesis,which effectively enhances the matching rate of person re-identification,and has positive scientific significance and application value.First,in order to tackle the deformation,rotation and translation of the target in the image,this thesis proposed a fusion method of color features and texture features,and the pyramid model is used to express the target pedestrian characteristics better.Combining the feature extraction with the distance measurement learning,we improved the re-identification rate more than 40%through the experiment analysis on VIPeR dataset.Second,combined with the advantages of convolutional neural networks in self-learning features,this thesis proposes the fusion of hand-crafted features and CNN features to achieve the performance complementarity between two characteristics,in order to build a more robust person re-identification model,along with applying the Softmax Loss function to the final layer of the network for multi-classification.To enhance the expression of feature fusion network model,network training is conducted on Market-1501 dataset.The trained network model is experimentally analyzed on the VIPeR dataset and CUHK dataset.The results show that the reidentification rate of the proposed algorithm can be improved to over 50%.Third,considering the peculiarities of local feature misalignment in pedestrian feature extraction,this thesis proposes to use full convolution network to extract pedestrian feature maps and construct pedestrian local feature extraction network.After regional detection,global pooling and linear dimension reduction,the local features are extracted to get a complete expression of pedestrian characteristics.Subsequently,the improved Triplet Loss function is used to measure the similarity.To verify the effectiveness of the proposed algorithm,an experimental analysis on the CUHK dataset shows that the re-identification rate of the proposed algorithm can reach more than 70%.In addition,considering the practical application of the scenes,this thesis uses just one convolutional neural network to build pedestrian detection network and person re-identification network for pedestrian detection and person re-identification.Then we builds person reidentification samples for experimental analysis.As a result,this thesis improves the reidentification rate to above 75%.On the basis,in view of the actual monitoring scenarios,a technology demonstration system of person re-identification is designed and the visual application of cross-camera person re-identification is preliminarily verified.Finally,this thesis summarizes the research content and work carried out,and looks forward to the future research direction.