Research Of Person Re-Identification Method Based On Robust Feature And Similarity Measurement

Person re-identification(ReID)is a crucial and challenging task of computer vision and also a focus in the field of intelligent video analysis.Its task is typically used for the person-matching task in a cross-camera scenario,which plays an important role in the work of suspect tracking,intelligent security and disaster warning.However,on the one in practical applications,due to the effects of low resolution,illumination changes,person posture,occlusion,etc.,there are large differences in the visual perception of person images,which makes the person re-identification process very difficult.On the other,robust similarity measurement is an important issue for person re-identification while most existing ReID models estimate the similarity between the Query images and the candidate Gallery images by computing the Euclidean distance between features,while ignoring the relationship of different samples,resulting in inaccurate similarity estimation.In view of the above problems,this dissertation proposes a feature extraction method of grayscale information enhancement based on bi-residual network and a novel similarity measurement method in the perspective of feature extraction and similarity measurement.The main content and innovation of this dissertation are as following two points:(1)Aiming at the phenomenon that only the single RGB images are used to extract person features in existing person re-identification algorithms,which causes the algorithms to be too sensitive to color information,a person feature extraction method of grayscale information enhancement is proposed,which is used to improve the robustness of the algorithm to color changes on existing datasets.Firstly in the proposed method,RGB images and corresponding grayscale images are taken as the input data of the bi-residual network to extract the RGB and grayscale features of person images respectively.Secondly,the two features are are concatenated in the channel direction to obtain the fusion feature,which increases the proportion of grayscale information,thereby weakening the color information and effectively reducing the network's sensitivity to color information.And then the fusion feature is learned by the strategy of horizontally evenly division.Finally,the RGB feature,grayscale feature and fusion feature are concatenated along the horizontal direction to obtain the robust combination feature for the testing process of person re-identification.In the training process,an independent loss function for each feature is adopted to optimize the model.Experiments show that the feature extraction method proposed in this dissertation can effectively solve the adverse effects of color similarity and color inconsistency on person re-identification task,and improve the mean average precision(mAP)and first hit accuracy rate(Top-1)of the algorithm.(2)Aiming at the problem of inaccurate similarity measurement in the existing person re-identification algorithms,a novel similarity measurement based on graph convolutional network(GCN)is proposed and named as PrGCN(Probability GCN),PrGCN regards the computing of similarity as a prediction problem of the link probability between node pairs and each person is regarded as an instance node.The greater the link probability,the more similar the two are.Firstly,construct Instance Centered Sub-graphs(ICS)around each instance node which depict the rich local context information.Secondly,input the ICS to a GCN to infer and predict the link probability of node pairs.Finally,a similarity ranking of the Query in Gallery images are gained according to the predicted probability.Experiments show that PrGCN improves the mAP and Top-1 accuracy of ReID significantly,yielding favorably comparable results to state-of-the-art methods.In addition,the proposed PrGCN can be easily embedded into other deep learning architectures instead of Euclidean distance metric while delivering significant performance improvements.