Person Re-identification Research Based On Visual Attention Mechanism

Public security is the precondition for the social stability and economic prosperity and development.It aims to prevent and control social crimes and various accidents,to protect people and their property.Nowadays,to escort the national public security,the intelligent monitoring system is further improved,whose main function is to analyze the monitoring data automatically.Person re-identification(re-id),which is seen as the one of the main applications of intelligent monitoring systems,has been a challenging research topic in the academic world.Its main job is that judging whether two people from different cameras with non-overlapping views are the same person.And based on this further other recognition tasks such as pedestrian tracking,motion trajectory prediction and so on.The key of person re-id is how to extract more discriminative representations from pedestrian images.Various challenges exist through out person re-id process,including background clutters,illumination variation,pose variation,occlusion,etc.Because of these challenges,person re-id has become the widespread research in the academic worldInspired by the way of human visual observation,this paper proposes a novel method based on deep convolutional neural network(CNN),which includes the visual attention mechanism to automatically learn human body feature.This paper aims at exploring the more discriminative person representations.On the other hand,the model compression algorithm is employed for further optimizing the proposed network.It reduces the parameters and speeds up under the premise of ensuring accuracy,which is beneficial to the application of person re-id.The main research contents of this paper include two aspects,visual attention mechanism model and model optimization:(1)The visual attention mechanism imitates the way of human visual observation.When an object is observed,human vision will ignore the global information or the background,and selectively focus the visual attention on local parts of the object.The key is to understand the global information via local parts information.For person recognition based on CNN,the visual attention mechanism model can be targeted to weaken the useless information or background interference,so that the CNN is more focused on the feature learning of human body.In this paper,the proposed visual attention adopts the pedestrian' mask images as the attention source.Via the designed attention CNN modules of different depth,the attention probability maps which includes the distribution of human body is obtained.In the backbone network,the attention map is used for guiding the final pedestrian representation extraction.The experiments confirm the proposed visual attention model is effective.It can obtain more discriminative person representations and reddress the problem of the background clutters effectively.(2)Both model compression and model simplification are used for the proposed visual-attention model.This paper uses the methods of structure pruning and equivalent substitution for model simplification.In addition,the model compression method is via channel pruning on the training model,which can reduce the parameter redundancy while ensuring the accuracy.Extensive experiments on benchmarks show it can reduce the storage occupation and computational complexity,and further optimize the proposed pedestrian re-id method.