Research On Key Technologies Of Large-scale Visual Surveillance Visualization

Since the new century,video surveillance is widely used in people's daily life,but the traditional camera shooting range is limited,resulting in a waste of monitoring resources,and the shooting scene is mostly two-dimensional image data of a single camera,lacking the visual monitoring effect of three-dimensional scene.Video enhanced virtual environment(AVE)combines real-time video with 3D virtual scene,which can not only maximize the consistency of time and space,but also maximize the application of video analysis results and get full understanding.It has broad development space and application value in government departments,transportation sites,social security and other aspects.The premise of fusing surveillance video image into 3D virtual scene is to know the pose parameters of real camera in 3D virtual scene.Manual annotation is a common method to solve the camera pose,but this method is not only time-consuming and laborious,but also difficult to apply when the camera needs to adjust the pose in real scene.Most cameras have the ability to automatically adjust their posture.Unless supported by the manufacturer's software,the attitude information of the camera after adjustment cannot be obtained in real time,which leads to the inaccuracy and real-time fusion of monitoring video and 3D virtual scene.In this thesis,based on the existing 3D virtual scene,the key technology of large-scale visual monitoring and visualization is researched by integrating real-time monitoring video and 3D virtual scene.The main work and achievements of the author are as follows:(1)This thesis analyzes the changes of camera parameters under special circumstances,and puts forward the method of real-time monitoring and real-time pose estimation.This method uses the color histogram of global features to quickly detect whether the camera parameters have changed.Then,msorb(multi-scale ORB)algorithm and fast GMS algorithm are proposed to obtain camera parameters in real time for accurate local feature extraction and matching.Msorb algorithm uses B-spline function to generate scale space to extract feature points in different scale space.It has fast operation speed and image rotation invariance,and improves the number of feature points and detection accuracy.In order to optimize the matching results and improve the matching speed,a fast GMS algorithm is proposed in the experimental study,that is,the four nearest neighbor five grid motion statistics scheme is used to optimize the matching,so as to ensure the real-time automatic fusion of the monitoring video and the 3D virtual scene after the camera parameters change,so as to avoid the manual complicated operation and lag.(2)When the camera pose parameters are known,this thesis analyzes the corresponding mapping region of 2D video image in 3D virtual scene,and adopts the improved shadow Map algorithm solves the problem caused by the inconsistency of visibility mapping between the perspective of surveillance video image and the three-dimensional virtual view,so as to determine the correct mapping area of two-dimensional video image in the three-dimensional virtual scene,and make the three-dimensional virtual scene get the correct monitoring video texture.For multi-channel cameras,a method of stitching and fusing multi-channel video and 3D scene model is proposed.On the basis of 3D rendering engine and GPU shader technology,the real-time fusion of real-time monitoring video image and 3D virtual scene pixel by pixel is realized.The experimental results show that the method is effective and practical,and realizes the visualization of large-scale visual monitoring,which is conducive to the monitoring personnel to analyze and process the monitoring video situation,and to enhance the application space of video virtual environment.