Person Re-identification Based On Adaptive Feature Clustering Network

With the continuous development of social information construction,the number of surveillance cameras and surveillance video data in the city is increasing sharply.Manually browsing and processing all surveillance videos has become time-consuming and laborious.Therefore,how to use computer to automatically integrate surveillance video information is important.Person re-identification technology can retrieve specific people from multiple camera video captures.This technology is widely used in intelligent video surveillance,intelligent transportation,intelligent security and other fields,but due to different video shooting angles,person movement posture changes,People background changes are complex,camera shooting quality is uneven,foggy and heavy rain and bad weather,so that the current person re-identification accuracy is not high,there is still a certain gap from the actual application.At present,the mainstream person re-identification algorithm consists of two parts: feature extraction and similarity measure.The former is used to extract the representative and robust person characteristics,and the latter projects the person features into a more distinguishable distribution space that improves the retrieval ability of the person re-identification model.In summary,this paper proposes three optimization methods for feature extraction and similarity measure:An adaptive feature clustering network structure is designed to extract person features with high representation ability and robustness.The adaptive feature clustering network consists of three parts: the basic feature extraction network,the feature clustering module and the classification module.The Res Net-50 network structure is selected as the basic feature extraction network.After obtaining the person feature map,the custom plane symmetric filling is used.A layer of convolutional layer classifies the feature maps by pixel level,and clusters the same features through maximum pooling to obtain the feature vectors.The classification module consists of a fully connected layer,a BN layer,a Re LU layer,and a fully connected layer.The person feature is projected onto the category label,and the sum of the cross entropy of the middle layer feature of the residual network and the category feature of the overall network end is taken as the final loss of the network,and the network multiplicity loss fusion is realized.Finally,the adaptive feature clustering network is proved to have a higher person re-identification accuracy by experiments.The metric loss function AC-Margin is proposed to replace the cross entropy loss in the original network.In order to make person features extracted by adaptive feature clustering network have better spatial distribution and improve the ability of the network to retrieve person images.The feature loss information is compressed in the metric loss function Asoftmax and the feature loss function Cosine Margin.AC-Margin combines the optimization features of both,taking into account angles and confidence.Spatial compression is performed to make the spatial feature distribution of people extracted by the network better,that is,to narrow the similar feature spacing of people and to expand non-similar person feature spacing.Experiments show that the increase in AC-Margin loss can significantly improve the retrieval rate of person images in the original network.Based on the modeling method of the attention mechanism in the CBAM module,the attention mechanism is used simultaneously for the channel dimension and the spatial dimension.The attention model is added to the basic feature extraction network and the feature clustering module respectively.At the same time,it has a feature clustering module and an attention mechanism module to improve the directionality of the network when extracting features,it is proved by experiments that this method makes the person features extracted by the network more representative and robust.