Research On Pedestrian Re-identification With Dense Multi View Field

With the gradual development of the national economy,people's life has entered a well-off society,and the requirements for social safety and property safety have also increased.At this time,the surveillance camera device will play a very important role.And for the large amount of pedestrian data that comes with it,it is obviously very difficult to find a specific pedestrian in the image data manually.Therefore,how to analyze the effective information under multiple cameras at the same time and accurately locate the target pedestrians in images has become a problem of great research value.At present,in the related research work of pedestrian re-identification,the pedestrian features extracted by Convolutional Neural Network have some deficiencies such as lack of information,which leads to weak discrimination of the obtained pedestrian features.Therefore,this paper proposes a Dense Multi View Field pedestrian re-identification model(DMVF),which uses the fusion of multi view field features to express pedestrian features.The work contents are as follows:(1)Firstly,the Baseline is proposed.The ResNet50 is used to extract features,and the metric learning method is selected to measure the similarity between pedestrians.(2)The densenet is applied to pedestrian re-identification instead of the original ResNet50 in Baseline as the basic feature extraction network,and use its advantages of feature reuse to retain more high-level detail features of lower layers;(3)The multi view field features are obtained by atrous convolution,and a multi view field structure which is more suitable for the model in this paper is proposed according to the atrous special pyramid pooling.(4)In order to integrate the multi view field structure into Baseline,a Dense Multi View Field(DMVF)method is proposed,which achieves the combine of the low-level semantic high-resolution features of the lower layer with the high-level semantic high-resolution features of the higher layer to obtain details.High semantic feature images of features and positioning capabilities.In order to prove the validity of the proposed algorithm,this paper chooses to compare three representative data sets in the pedestrian recognition research:Market-1501,DukeMTMC-ReID and CUHK03 data sets for experimental verification,and for each data set selects several other popular pedestrian re-identification algorithms to compare with.The experimental results show that the model using DenseNet as the feature extraction network and the Dense Multi View Field has improved in accuracy compared to Baseline.The dense multi-view field model proposed in this paper is better than other mainstream algorithms,and has achieved better pedestrian recognition effect,effectively solving the problem of insufficient feature extraction in pedestrian re-identification.and has a strong robustness.