Research And Application Of Pedestrian Detection And Tracking Algorithm In Video Surveillance Scenes

With the development of deep learning technology and related algorithms,intelligent video surveillance has played a huge role in security,remote command,emergency rescue and other fields.However,for the detection and tracking of pedestrians,such as objects with changeable characteristics and uncertain moving directions,the accuracy and stability of current intelligent video surveillance cannot meet the actual requirements of the surveillance system for accuracy and response speed.Therefore,starting from intelligent video surveillance,this paper conducts in-depth research and improvement on pedestrian detection and tracking algorithms,and realizes an efficient pedestrian detection and tracking monitoring system.The main tasks completed are as follows:(1)Firstly,aiming at the problem of inaccurate detection position caused by partial occlusion and non-rigid body changes in the process of pedestrian detection,the application of occlusion Data Augmentation and Deformable Convolution to the Center Net network is studied to improve the detection effect.Secondly,according to the specific pedestrian feature distribution and the problem of mutual occlusion between pedestrians during the detection process,the Gaussian kernel distribution in the original model is improved to make it more suitable for the detection results of pedestrians,thereby the missed detection caused by the mutual shielding of pedestrians is reduced in video.Finally,the results are evaluated.The experimental results show that under the condition that the model can be detected in real time,the Average Precision and stability of the detection results have been significantly improved.(2)Data association is carried out on the basis of detection.Firstly,in order to solve the problem of inaccurate appearance similarity estimation caused by the unfixed position of pedestrians in the detection frame,an attention mechanism is used to extract appearance features.Secondly,a dynamic weighted fusion strategy is proposed based on appearance characteristics,motion characteristics,and geometric characteristics for reduced accuracy and frequent identity switch caused by pedestrian missed detection and incorrect matching.Then,for the trajectory separation caused by the long-term missed detection of pedestrians and the same identity trajectory in the adjacent video,this paper proposes a trajectory connection method,which effectively reduces the trajectory separation caused by long-term occlusion and the trajectory connection across the video.Finally,experiments are carried out on the data set MOT16,and the results show that the tracker implemented by this method has a significant improvement in the performance of MOTA,MT,ID Switch and other metrics.(3)Aiming at the problems of data redundancy,inflexible query and analysis in traditional video surveillance,this paper implements a pedestrian detection and tracking system in video surveillance.The system collects data through the camera,calculates and stores the collected video through the server,and the local client realizes the receiving and loading of the calculation results,the visualization of the data,and the interaction with user.Finally,through the data communication collaboration between the camera,server,client,the functions of result storage,reduction of data redundancy,and visual analysis in video surveillance scene are realized.