Research And Implementation Of Person Detection And Re-identification Based On Deep Learning

In recent years,the research on artificial intelligence technology is particularly important with the rapid development of information and intelligence.As an important part of artificial intelligence technology,computer vision has a wide range of application value in the fields of intelligent monitoring,driverless and intelligent transportation.Domestic and foreign scholars have done a lot of research on person,because people are regarded as one of the most noticeable objects in vision.In order to achieve the task of person reidentification in the monitoring scene,improve the speed of person detection and the accuracy of person reidentification,this paper studies the problem of person detection and re-identification in the intelligent monitoring,the main research contents are as follows:(1)In order to detect the person in the monitoring image quickly and accurately,a lightweight person detection model is designed,which is introduced the RFB(Receptive Field Block,RFB).Specifically,in order to reduce the number of model parameters and improve the speed of person detection,a lightweight feature extraction network is designed to extract the effective features in the image whose maximum number of convolution channels is 256.In addition,in order to improve the accuracy of model,the improved RFB is introduced into the lightweight feature extraction network to extract features with stronger discrimination.Finally,in order to detect the person with different sizes in the image accurately,6 feature maps with different sizes are selected from the feature extraction network,and different anchor are designed for each feature map according to the characteristics of person shape distribution to carry out category prediction and location prediction.The experimental results show that the model can not only detect the person in the image accurately,but also improve the speed of the existing object detection model.(2)In order to solve the matching problem between person images,a person re-identification model with multiple granularities is designed.Specifically,in order to extract person features with more comprehensive information and stronger discrimination,a multiple granularities feature extraction network is designed,which extracts global and local features of different granularities from the feature maps of Conv4 and Conv5 of Res Net50 backbone network.In addition,in order to learn better network parameters,this model joint representation learning and metric learning,and set Label-Smoothing ID loss and Batch-Hard triplet loss in the network.Experimental results show that the model can accurately match the relationship between person images,and has certain effectiveness and advance.(3)In order to realize the task of person re-identification in the monitoring scene,a person detection and re-identification system is designed and implemented.Specifically,this system first detects all person from the monitoring images using the lightweight person detection model with RFB.And then,it extracts the features of the query image and the detected person image using the multiple granularities person reidentification model,and calculates the Euclidean distance between the features.Finally,this system finds the person whose ID is same as query's ID and marks this person.The experimental results show that this system is effective and practical.