Research On Multi-level And Multi-scale Pedestrian Re-identification Algorithm

With the continuous improvement of the quality of monitoring equipment and the continuous decrease of the cost,surveillance video,as an important means to maintain social public security,has attracted extensive attention in recent years.With the outbreak of COVID-19,the epidemiological investigation of confirmed or suspected cases is very important.At present,it takes a lot of manpower and material resources to find the moving track of cases from a large number of videos,and person re-identification(person re ID)is a good alternative technology.However,the background in public places is complex,and a series of problems,such as light,occlusion,and angle of view,make person re ID still a great challenge.Therefore,extracting features with discriminant ability and robustness from images is still a key problem in current research.With an eye on the complementary relationship between global features and local features,this paper uses attention mechanism,human body semantic parsing,and feature selection to improve person re ID based on deep learning and improve the feature representation ability of the deep convolution model.The main research contents of this paper are as follows:(1)We propose a hybrid-attention guided network with multiple resolution features for person re-ID,which aims to enhance the representation capacity of the CNN models to discriminatively learn the pedestrian features.First,the model can fuse the deep and shallow features to solve the problem of information loss in high-level feature maps.Then,the fused features are segmented horizontally to obtain multi-resolution global features and local features to alleviate the effect caused by the misaligned bounding boxes.Besides,we also introduce two different attention mechanisms to mine the salient regions.Through the joint action of these steps,the feature representation ability of the model is greatly improved.Extensive experiments present the superiority of our model on four large datasets.(2)Existing person re ID methods adopt stripes or blocks to obtain the local features.However,some stripes or blocks may not contain pedestrian information or contain less information.If these stripes are used for person re ID,the representation ability of the model is reduced to some extent.Besides,the stripes cannot solve the local misalignment problem.For these,this paper proposes a local-aligned person re ID algorithm based on human parsing and feature selection,which uses global features,local features of stripes,and local features of human body parts to improve the recognition rate of the model.First,a novel Salient Feature Selection unit was designed to suppress the stripe containing less pedestrian information and enhance the stripe containing more pedestrian information.Secondly,we designed a new attention-based human parsing feature learning unit to obtain pure local features of human body parts and to solve the problem of misalignment.Extensive experiments show that the proposed method is superior to the existing methods.