Research On Pedestrian Reidentification Method Based On Deep Learning

In recent years,more and more attention has been paid to security issues,and monitoring equipment is becoming more and more common.It takes a lot of manpower to find target information from massive video.Pedestrian recognition is a technology that uses computer vision technology to determine whether there is a specific pedestrian in the image or video sequence.The front-end device detects the face and pedestrian image in real time,and completes the task of pedestrian recognition through face recognition and pedestrian recognition technology.Pedestrian recognition technology plays an important role in the field of intelligent security.In the surveillance video,because of the lack of computing power of the front-end camera,it is unable to detect human faces and pedestrians in real time;because of the camera resolution and angle,it is unable to obtain high-quality human faces,and the pedestrian cross camera mutation is relatively large,which is the key to restrict the development of pedestrian recognition technology.To solve these problems,many scholars compress the depth model to meet the real-time requirements;design a better loss function to learn a better face recognition network;and use the measurement method to complete the pedestrian recognition task.In this paper,we study how to improve the accuracy of pedestrian recognition and the real-time performance of detection model.The main research contents are as follows:(1)Design and implementation of face quality evaluation algorithm modelAiming at the problem that low quality face image affects recognition speed and recognition accuracy in face recognition process,a face quality evaluation method based on deep convolution neural network is proposed.Firstly,regression and classification methods are used to predict the probability of eight attributes related to face quality,such as yaw,pitch,norm value,mouth opening,occlusion,blur,dim,eyes closing,etc.;secondly,branch weight W is obtained based on the optimization of video pass rate objective function;finally,each branch is weighted to obtain the face quality score;the higher the score,the better the quality of the face image.The face image that does not meet the quality score will be filtered out,and the face that meets the quality will be recommended to the face recognition model,so as to improve the face recognition accuracy and optimize the face recognition process.(2)Pedestrian recognition based on pedestrian attributesAiming at the low accuracy of pedestrian recognition.A multi task network structure is designed to predict pedestrian attributes by multi-scale,and the feature vector of pedestrian attributes is fused to the pedestrian identification branch through a repnet network,so that the pedestrian identification branch contains more fine-grained characteristics of pedestrians,thus improving the accuracy of pedestrian identification.(3)Model compression algorithm based on characteristic distillationIn order to solve the problem that depth model can not be deployed in the front-end low-power equipment,this paper uses knowledge distillation method to compress the detection model,so that the model can maintain the accuracy and speed up the operation of the model.First,train the network of teachers' retianet person and retianet face;then,the network of teachers guides the network of students' mobilenetv2 person and mobilenetv2 face to complete the training,helps students learn more hidden information online,and makes the network of students improve speed while maintaining accuracy.The performance of the research method in the test set is verified.First of all,the performance of face quality analysis model is tested on the test set.The experimental results show that the face quality model can filter out low-quality faces in real time.Under the same hardware environment,the speed of the whole face recognition process is increased by 20 times,and the accuracy is increased by 2.3%.The experimental results verify the effectiveness of the face quality model.Then,the validity of pedestrian recognition algorithm is verified in the open data set market1510.MPAP is 30.27% higher than that before optimization.Compared with the current optimal DSC algorithm,m PAP is 5.72% higher.The experimental results show that the algorithm can effectively improve the accuracy of pedestrian recognition.Finally,the performance of the compressed detection algorithm is verified on the open datasets of widerface and INRIA.The experimental results show that the model size is compressed more than 100 times by using the knowledge distillation method,and the map is only reduced by a few percentage points,which verifies the effectiveness of the algorithm.