Research On Pedestrian Identification Based On Deep Learning

As more and more monitoring devices are deployed in various cities around the world,the technology of computer intelligent analysis and processing of video image data is becoming more and more mature.As an important application direction in intelligent video analysis,the pedestrian identification mainly extracts and processes the video frame sequences of pedestrians under different cameras to detect the moving objects in the cameras,and then judges whether they are the same target pedestrians according to the characteristics of pedestrian objects(such as face,clothing color,etc).However,the pedestrian identification in the actual environment is easily affected by lighting,pedestrian posture and shooting angle and other factors.The pedestrian images obtained by different cameras are also very different,which seriously affects the accuracy of pedestrian identification.In this thesis,the surveillance video in the building is used as the research object,how to identify pedestrians based on deep learning algorithm and overcome the existing problems of pedestrian identification and improve the speed and accuracy of pedestrian identification.The research contents of this thesis are as follows:In this thesis,the pedestrian detection based on traditional methods and deep learning algorithms is studied and compared,and the YOLO target detection framework is determined as the basis of our method.In order to solve the problem that pedestrian detection and location is not accurate enough due to small target detection and pedestrian occlusion,a pedestrian detection algorithm based on the combination of traditional ViBe and YOLO is adopted to realize the pedestrian detection of internal personnel in surveillance video.Firstly,ViBe algorithm is used to detect pedestrians at first stage,and some pedestrian frames are selected.Then,the pedestrian frames are sent to the YOLO network for secondary detection.The second pedestrian detection based on deep learning uses K-means algorithm to complete the clustering of prior frames,and then uses CSPDark Net53 to extract pedestrian features.The experimental results show that the detection algorithm based on the combination of ViBe and YOLO optimizes the regression of pedestrian boundary box,improves the process of non maximum suppression,and the positioning accuracy of pedestrians,and lays a good foundation for the next pedestrian identification.In order to solve the problem that the apparent difference of pedestrians is small and the external characteristics are easily affected by the environment,which makes it difficult to confirm the identity information of pedestrians,On the basis of detecting the pedestrian target,the pedestrian face is taken as the research object to complete the pedestrian identification based on face recognition.The proposed pedestrian identification process includes four steps:face detection,face feature extraction,face feature matching and identity authentication.First of all,the face detection process is based on the YOLO model.According to the learning of face prior knowledge,this thesis selects and reduces the size of the network prediction box by using the proportion knowledge of face during the initialization of network prediction,and determines the size of the prior boundary box again for face detection.At the same time,the loss function of face detection is improved by introducing the idea of Focal Loss to solve the problem of unbalanced samples in the face data set.Secondly,the residual network is constructed to complete the face feature extraction model,and the Euclidean distance is selected to match the recognized face with the face in the local face database.If it is confirmed that it is a local person,the system will automatically output the identity information of the person.The experimental results show our algorithm based on deep learning can accurately extract face features and effectively complete the task of pedestrian identification.Finally,through the overall design of pedestrian identification system,a concise GUI interface is designed to complete the function of pedestrian identification.