Multi-view Feature And Strong Constraint Dissimilarity Space Cross-domain Pedestrian Re-identification

In recent years,with completing large-scale projects such as "Smart City","Safe City" and " Snow Bright Project ",China’s video surveillance network has become the largest in the world.Much video surveillance has a large amount of video data.How to analyze these data quickly has become a key problem to be solved.Searching with eyes is time-consuming,hard-working and error-prone.Therefore,developing intelligent video surveillance is imminent.Pedestrian re-identification(Re-ID)is a new technology in intelligent video surveillance analysis.Its purpose is to retrieve target pedestrians in non-overlapping shooting areas,and it has become a research hotspot in computer vision.Through introducing attention mechanism,image feature segmentation,and triple loss,problems such as low resolution,posture change,viewing angle difference,illumination change,occlusion,complex background,and clothing difference have been overcome,and pedestrian re-identification has been basically solved in the limited domain space.The accuracy of these Re-ID methods is over 95% on public datasets such as Market1501 and Duke MTMC-Re ID,etc.However,there are still many problems in open space.The difficulty in open space is the pedestrian images involved in the model training and test cannot cover all situations in open space.Scholars have begun cross-domain Re-ID research to address spatial differences.Among the existing methods for cross-domain Re-ID,the accuracy of the domain adaptive method on Market1501 is only 58.2%,and the accuracy of the method based on transfer learning on Market1501 is 90%.It can be seen there is still room for improvement in the cross-domain Re-ID.This thesis studies cross-domain Re-ID methods from perspectives such as feature design,metric space improvement,and feature secondary encoding.The main contents and innovations are as follows:(1)To solve the problem that pedestrian features are limited by cross-domain,a multi-view feature cross-domain pedestrian re-identification model is proposed.Viewing angle change is one of the most important reasons affecting the accuracy of Re-ID.Based on View-dependent Frame of Reference in Cognitive Neuroscience and domain generalization,a multi-view feature cross-domain Re-ID model is proposed.Divide the viewing angle into three standard angles of view: front,side and back,and build a multi-view Re-ID dataset.This model is designed based on Capsule Network,and a viewing angle classification loss and a pedestrian ID loss are used to capture the features needed for view discrimination and to constrain consistency of pedestrian identity.It can reproduce human cognitive behavior and make a view-related but domain-independent feature for cross-domain Re-ID.Experimental results show that it can filter out the effective information such as limbs,head and face from pedestrian images,remove the background interference,and do cross-domain pedestrian re-identification.(2)To solve the problem the model occupies many resources,multi-view feature cross-domain pedestrian re-identification model compression methods are proposed.This cross-domain Re-ID based on Capsule Network shows good performance.However,when the input space dimension is large,Capsule Network cannot effectively reduce the feature dimension and consumes many computing resources.Also,the coupling coefficient of the dynamic routing algorithm in Capsule Network has a minimization trend,which hinders gradient backpropagation.To solve these drawbacks,two improvement schemes are proposed.One is a deep multi-view feature cross-domain Re-ID,which integrates the attention mechanism in the convolution feature layer,increases the number of layers,and redesigns the coupling coefficient generation function of the dynamic routing algorithm in the digital capsule layer.The second is a big-little multi-view feature cross-domain Re-ID.A lightweight primary capsule layer is designed,which uses two kinds of capsules with different receptive fields to reduce the feature dimension.The coupling coefficient generation function is further optimized to improve the practicability of this cross-domain Re-ID model.(3)To solve the problem the accuracy of the domain generalization model is not high,a cross-domain pedestrian re-identification model based on transfer learning with strong constraints is proposed.The domain generalization Re-ID models have good domain adaptability,but the accuracy is not high.The main problem is the common feature space between different identification domains is not well constructed.A strong constraint dissimilarity space for cross-domain Re-ID model based on transfer learning is proposed.It transfers the source domain knowledge to the target domain,builds a common feature space effectively for the source and target domain samples,and solves the problem the dissimilarity space made by the DMMD model has weak constraints on the sample distribution.In addition,an appearance spatiotemporal loss to enhance feature discrimination and a domain adaptation classification boundary optimization method suitable for transfer learning are also proposed.The methods in this chapter can greatly improve the accuracy in the target domain of the cross-domain Re-ID algorithm.(4)To solve the problem the pedestrian image features generated by existing Re-ID models are not fully utilized,a multi-task contrastive learning adversarial autoencoder is proposed.Rerank is the method of reordering Re-ID results,which can improve the accuracy based on the Re-ID results again,showing the features generated by the existing Re-ID models have not yet been fully utilized.These models have room for improvement.To solve this problem,combining the idea of Rerank and multi-attribute Re-ID,a multi-task contrastive learning adversarial autoencoder is designed to perform secondary encoding on the generated Re-ID features based on transfer learning.First,make full use of the DG-Net++ model to transfer the pedestrian appearance features of the source domain to the target domain,and generate the color attribute labels of pedestrian clothes in the target domain.Secondly,use the DBSCAN to generate the ID pseudo-labels of images in the target domain.Finally,using these two kinds of labels,an improved adversarial autoencoder is designed based on multi-task contrastive learning,and the features generated by existing cross-domain Re-ID models,including the strong constraint dissimilarity space cross-domain Re-ID model and DG-NET++,are re-encoded to improve the model accuracy.