Research On Enhanced Deep Feature Based Person Re-identification Technology

Person re-identification is an important technology in the field of computer vision,which aims to match pedestrians across different surveillance camera views.Feature representation and metric learning are two critical components in person re-identification technique.Feature representation is a challenging problem because the visual appearance of pedestrians can change drastically across views.Therefore,it is critical to design a discriminative and robust feature for representing pedestrian images to face the challenges in person re-identification task.In order to deal with the problems in person re-identification,this paper propose an enhanced deep feature based person re-identification method.First,the method takes identification-verification network as basic frame and introduce a multi-level attention mechanism in it for extracting discriminative deep feature.Specifically,the regional-level attention mechanism is able to locate key regions of the pedestrian image to cope with misalignment problems caused by detection errors.Channel-level attention and pixel-level attention can calibrate the feature maps at the channel level and pixel level,respectively,making the network extract more accurate deep features of pedestrian.Then,the method implants hand-crafted feature into the network by making hand-crafted features participate in the construction of network and the generation of the final features,in order to utilize the complementarity between the manual features and the convolutional neural network features.Finally,this paper designs a feature reconstruction module to learn an efficient way to fuse hand-crafted features and deep features.The enhanced deep features generated in this way will have the advantages of both types of features.In this paper,extensive experiments and comparative analysis are carried out on the existing person re-identification datasets,including Market-1501,CUHK03 and VIPeR.Experimental results show that the method proposed in this paper is superior to the baseline in Rank-i accuracy,cumulative matching characteristic curve and mean average precision.The method is competitive with a wide variety of state-of-the-art methods,which prove the effectiveness of the proposed method.