Research And Implementation Of Person Re-ID Based On Semantic Segmentation

Person Re-ID is a technology that efficiently retrieves the same pedestrian information on different shots of pedestrian pictures(videos).It is also a basic technology in the security field.It can perform a fugitive hunting,trajectory tracking and other tasks at low cost and high accuracy.However,due to the differences in camera angle,shooting time,and human posture,it has brought many obstacles to the application of the technology.More and more models focus on the combination of local features and global features to solve problems encountered in the field.Semantic segmentation classifies each pixel of the picture and can be used to extract local features of pedestrians.Therefore,it is of great research value to use the pedestrian semantic segmentation results as a bridge connecting pedestrian global and local features.Aiming at the above problems,this thesis uses semantically segmented information to assist pedestrian re-retrieval tasks.The main research results include two aspects.First,in the field of semantic segmentation,a semantic segmentation network based on context information is designed.In this network,the advantages of two models of deeplab v3 + and EncNet are combined.In addition,part encoding loss is proposed to help the semantic segmentation network to recognize small objects,and part encoding attention is used to encode the context information of the picture by deploying the attention mechanism to improve the performance of model semantic segmentation.mIOU was obtained on the VOC2012 dataset.= 83.56 % score.Secondly,in the field of pedestrian re-retrieval,a pedestrian re-retrieval network based on semantic segmentation assistance is designed.The network uses a multi-branch structure,which combines stripe partitioning and softening of sub-information using semantic segmentation,which significantly improves the performance of the model.It has achieved mAP = 87.66 and P1 = 95.73 on Market1501.In addition,these two models were also applied to Hikvision's internal industrial-level data sets,and the model performance was excellent.