Multi Granularity Person Re-Identfication Based On Local Attention Mechanism

In recent years,with the development of intelligent monitoring system and intelligent image processing technology,person re-identification technology has a wide range of application prospects in the fields of suspect tracking,intelligent human search,self-service supermarket and intelligent robot and so on.Due to the complexity of the real scene,person re-identification still faces many difficulties in practical application,mainly in the changeable pedestrian attributes such as posture,gait and clothing,and the serious interference of environmental factors such as illumination change,camera angle difference and object occlusion.These differences make it very difficult to extract robust feature representation of human body.In this thesis,a multi granularity person re-identification network combined with local attention mechanism is proposed to extract the diversity features with stronger identification ability in complex pedestrian scenes.At First,considering that the spatial attention algorithm in the existing pedestrian re recognition research is easy to over fit the external features such as clothing and ornaments,which are replaceable and not universal,this thesis proposes a local attention method and constructs a complete local feature extraction network based on it.This method can generate a series of ROI maps related to human parts through convolution network without additional data annotation,and then obtain the local feature representation of human parts,so that the network can fully learn fine-grained local features.Secondly,based on local attention algorithm,this thesis proposes BA-Drop Block and an attention feature regularization mechanism.BA-Drop Block overcomes the randomness and blindness of the existing mainstream convolutional network regularization methods by making the network mining more universal and discriminative features under the guidance of attention selection mechanism.The attention feature regularization mechanism constructs a category center for each local feature,and enhances the training effect of the model by learning the key regions of the constraint network.Thirdly,considering that the features of pedestrians in real scene change greatly and the interference of environmental noise is serious,this thesis proposes a multi granularity model,LAG-Net,which combines the global features of coarse-grained image and the local features of fine-grained human body.This network can not only considers the coarse-grained cues captured by a generalized part-based structure but also extracts semantically similar part cues through a local attention structure.In order to verify the effectiveness of the above method,experiments are carried out on three mainstream public data sets: Market-1501,Duke MTMC-re ID and CUHK03.Experimental results demonstrate that our proposed method achieves state-of-the-art results on all datasets.Especially in the CUHK03 dataset,we achieve a 5.3 % performance gain on accuracy over the best existing approache.