Traffic Incidents Construction Based On Augmented Reality

The technology of traffic video incident detection judges the abnormal incident such as abandoned objects,parking,and pedestrians in the video through automatic analysis and understanding of the traffic video content,so as to issue warnings in time to avoid traffic accidents.For the research method of this technology,one of the most reliable technical routes is to obtain samples of a large number of traffic incident,and analyze these samples to establish characteristic models of different incidents.However,some traffic incidents do not occur very often,making the samples acquisition of these traffic incidents very difficult.The serious lack of samples limits the development of the technology of traffic video incident detection to a certain extent.In this thesis,a method for traffic video content augmentation is proposed.Using augmented reality,the road plane is innovatively used as the "link" of the virtual-real fusion.The virtual model is superimposed on the pavement area of the real traffic scene video without any calibration object,thus constructing realistic traffic incident videos.The main research contents of this thesis are:(1)Research and implementation of geometric consistency in virtual-real fusion.Firstly,the camera calibration is based on a single image using double vanishing point method,and the internal and external parameters of the camera and the geometric information of the road plane in the traffic scene are obtained.Secondly,the Blender modeling principle is studied to complete the virtual road plane and the virtual model of the OBJ format.Then,the coordinate transformation relationship between the virtual road plane and the real road scene is obtained by using the 3D reconstruction geometric transformation basis to ensure the position and size of the virtual model to be registered to the real scene.Finally,the occlusion problem in the augmented reality is studied.The background difference method is combined with the pixel-based traversal method to achieve accurate occlusion judgment,and the template caching technology is used to eliminate the objects occlusion.(2)Research and implementation of illumination consistency in virtual and real fusion.Firstly,the influencing factors of traffic scene illumination are studied,and the Lambert illumination model is established.The illumination parameters are solved by offline interaction to ensure the correct brightness and darkness information on each side of the virtual model.Secondly,the shadow rendering algorithm is studied and the shortcoming of the traditional shadow mapping algorithm using only one depth texture for calculations is found.An improvement to the traditional shadow mapping algorithm is proposed,and different types of depth textures are used for rendering static objects and dynamic objects.The experimental results show that the rendering speed of the algorithm is significantly higher than that of the traditional algorithm,which can efficiently generate shadows for virtual models in complex traffic scenarios.Using the method of this thesis,the content of traffic monitoring video of a certain section of Xianxun Highway has been augmented,and three types of traffic incidents that are difficult to obtain are constructed: high-speed pedestrian,parking,and abandoned object.From the experimentally generated video,the newly added dynamic and static virtual models meet the geometric consistency and illumination consistency requirements of the real traffic scene.The newly generated event video has a high degree of realism and can be used for traffic video event detection technology.