Research On Pedestrian Re-Identification Method Based On Deep Learning

At present,there are great safety risks in public places.For example,the news often reports that old people or children are missing in large amusement places.Pedestrian re-identification technology can effectively solve the problem of missing persons.It can quickly retrieve the missing pedestrians from the images taken by different cameras in the same area.However,in reality,pedestrian body parts will be affected by conditions such as different light,chaotic background,resolution level,different camera angles and occlusion,which will lead to interference in pedestrian re-identification model when extracting pedestrian features.There is no way to meet the requirements of the application scenario.Aiming at the above problems,this paper studies the pedestrian re-identification technology,with the main contents as follows:(1)In view of the influence of illumination,complex background,resolution,camera perspective and other factors,this paper proposes a pedestrian re-identification method based on feature fusion of joint channel and self-attention.The baseline model of this paper is ResNet50-IBN network.First,the 3x3 convolution in part of the residual blocks is replaced by multiple self-attention modules,so that the network can reduce the stacking of multiple convolutional blocks while modeling long-distance dependence,which is helpful to improve its own performance.Secondly,the channel attention mechanism ECA module is added to the ResNet50-IBN backbone network to make the network pay more attention to the feature information of the pedestrian foreground.Finally,the output features of the last three groups of convolutional blocks are fused,so that the last features extracted by the network include intermediate features and advanced features,so as to avoid the loss of key information that effectively distinguishes different pedestrians.(2)In view of the impact of occlusion noise on pedestrian images,this paper proposes an occlusion pedestrian re-identification method combined with human posture estimation to accurately locate human bodies.Firstly,considering that the feature fusion network of joint joint channel and self-attention has poor performance in the extraction of global features containing occlusive pedestrians,a residual shrinkage module is introduced for occlusion processing.Secondly,the human posture estimation model is introduced to locate the non-occluding area of the human body to extract local features of pedestrians.While effectively locating the local key points of the human body,partial occluding noise is abandoned.Finally,multiply the pedestrian features extracted from the baseline model with the pedestrian features extracted from the human body pose estimation network to obtain the global features and multiple local features.Considering that the testing and training between the local features during feature extraction will lead to the loss of the features in the non-occloccled part,the correlation relation module is used to make the global features and local features interrelate.It makes the model more capable of feature discrimination.