Research And Implemention Of Pedestrian Re-recognition In Complex Scenes

Pedestrian recognition in complex scenes is a very widely used algorithm in automatic driving,intelligent transportation,intelligent security and other fields.It is also a hot issue in the field of machine vision application.The core problem of pedestrian re-recognition is to realize cross-camera recognition and search for pedestrians in complex scenes.At present,the method for pedestrian re-recognition task has achieved good recognition effect,but there are still the following problems: On the one hand,the pedestrian attribute recognition is pedestrian recognition of other important basis,can improve the identification of pedestrians,the pedestrian attribute recognition in practical application,often attribute unbalanced situation of the training sample,influence the effect of the algorithm,especially when attributes can balance in severe cases,a network model of performance degradation is obvious;On the other hand,there is still room for improvement in the online effective mining ability of difficult positive and negative sample pairs in the network training of pedestrian rerecognition.Therefore,this paper has completed the following work and innovation for the above problems:First,a pedestrian imbalance attribute recognition method based on Res Ne St and hardsample online mining is proposed.The proposed method uses Res Ne St50 backbone network and channel self-attention module(CAM)to extract attention features with certain significance,and improves the performance of the model through label smoothing strategy,online mining method of pedestrian attribute difficulty samples(OHEM)and multi-attribute joint training method.Experimental results on RAP and PA100 K datasets show that the proposed method is superior to existing pedestrian attribute recognition models in overall performance,and has good applicability and strong competitiveness.Second,an improved multi-scale difficult triplet loss function based pedestrian rerecognition method is proposed.The method using Res Ne St50 as backbone network to extract the features,use a label smooth regularization of the cross-entropy loss function and improved multi-scale difficult triples loss function to improve the performance of the model,and introduces the loss balance weighted strategies to improve the multi-scale difficult triples loss function with cross entropy loss function for a joint optimization.Experimental results on market-1501 and Dukemt MC-Reid data sets show that the proposed method has certain advantages over existing pedestrian re-recognition methods in terms of overall performance.Thirdly,based on the above algorithm model and classical target detection algorithm,a pedestrian re-recognition system is constructed in complex scenes,which can be used to recognize specific pedestrian targets under different cameras.At the same time,a user interface is designed,which is simple and easy to operate.