Study On Person Re-Identification Based On Deep Learning

Person re-identification(ReID)aims to use computer vision technique to identify target person in images or videos across different cameras.It is also regarded as a subtask of the image retrival.Person ReID is widely used in criminal investigation,smart supermarket and multi-target tracking.Recently,deep learning has achieved great success in many computer vision tasks,person ReID based on deep learning is widely used in public safety,criminal investigation,and smart city,etc.Person ReID has different critical problems to be studied in non-occluded and occluded situations,respectively.For the standard person ReID in non-occulded situation,the critical problem is how to extract robust and discriminative features.For the obscured person ReID in occluded situation,the critical problem is how to detect and match the non-occluded areas shared by the partial and holistic images(occluded and non-occluded images are named as partial and holistic images in this thesis,respectively).According to whether the occlusion is pre-removed during the matching process,obscured person ReID is divided into partial person ReID and occluded person ReID.Based on deep learning,this thesis proposes some methods for the standard person ReID in non-occulded situation and the obstructed person ReID in occluded situation.The main contributions are concluded as follow:1.For the standard person ReID in non-occulded situation,the thesis proposes Aligne-dReID++framework based on Dynamically Matching Local Information(DMLI)method.DMLI first extract the local features of two images,and then compute the shortest path distance of it to can automaticly align the local features of two person images.DMLI can improve the robust-ness to pose variances.In addition,AlignedReID++is jointly learned with global features and local feature based on DML1.AlignedReID++can resolve the pose variances while keeping high efficiency.2.For the standard person ReID in non-occulded situation,the thesis proposes a strong baseline,which designs a BNNeck structure to resolve the inconsistance between classification loss and metric loss.With some widely used training tricks,the strong baseline only uses global features to achieve significant performance and efficiency.In addition,it exploits center loss to improve the intra-class compactness and inter-class separability of the feature.The proposed baseline is able to improve the performance of existing methods,so it can be considered as a research baseline for both the academia and industry.3.For the partial person ReID in occluded situation,the thesis proposes STNReID frame-work based on spatial transform networks(STN).For the partial person ReID,the occusions in the partial image are pre-removed.STNReID includes STN module and ReID module.STN module samples a patch from the holistic image to match the partial image by predicting a set of affine transformation parameters,and ReID module extracts global features to measure sim-ilarity of two images.In addition,a two-stage training manner is also proposed to improve the performance of STNReID.4.For the occluded person ReID in occluded situation,the thesis proposes Cross-Correlation Based Attention Pooling(CAP)method based on cross-correlation matching.For the occluded person ReID,the occusions in the partial image are not pre-removed.Applying cross-correlation operation to perform cross-convolution on deep feature maps of partial and holistic images,CAP method can get an attention map that focuses on the region shared between partial and holistic images.By giving higher weights to the local features of the non-occluded parts,CAP weakens the influence of occlusions to improve the recognition accuracy of the model.The research works on standard person ReID in the thesis have been widely used in the industry.In addition,the thesis conducts systematic and frontier research on obscured person ReID.