3D Reconstruction For Vehicle Based On Vedio Data

The three-dimensional reconstruction of the moving vehicles in the traffic monitoring video is not only conducive to vehicle identification and classification,but also to further enrich the content and form of traffic data.At present,the objects reconstructed based on image sequences are still objects or large-scale scenes.Therefore,it is of great significance to study the 3D reconstruction technology of moving vehicles in traffic video.Starting with 3D reconstruction,this paper consults a large number of domestic and foreign documents,analyzes and summarizes the development of the main steps in 3D reconstruction,and divides the 3D reconstruction of moving vehicles in traffic video into three main steps.The details are as follows:The first step is the segmentation of moving vehicles in the video.Scenes in traffic videos are often complex.In order to eliminate the influence of background and other objects on the reconstruction of the target vehicle,the vehicle needs to be extracted from the background.The principle and process of the classic Vi Be moving target segmentation algorithm are studied.For dealing with the problem that the fixed threshold foreground judgment method is prone to misjudgment and lead to target holes,adaptive gray threshold and color distortion threshold are used to replace fixed thresholds when determining whether a pixel is background or foreground,which reduces a large number of foreground holes.Aiming at the problem that the background model is prone to ghosts when initialized with neighboring pixels,ghost detection methods based on the histogram’s Tanimoto coefficient and Euclidean distance is brought in。Experiments show that the improved algorithm reduces ghost effectively。The second step is extraction and matching of features.The purpose of 3D reconstruction is to obtain the 3D coordinates of the target feature points,so it is necessary to extract and match the features of the reconstructed target.The principles and processes of the classic Scale-invariable Feature Transform(SIFT)algorithm and Random Sample Consensus(RANSAC)algorithm are studied.For the SIFT algorithm that uses all scale spaces for matching strategies,it is prone to the problem of low efficiency.A matching method of choosing the scale space suitable for the target object is proposed,which improves the efficiency of the algorithm while retaining more correct matches in the experiment.Aiming at the low efficiency problem caused by the random sampling of the RANSAC algorithm,the quality function is used to construct a semi-random sampling method of hypothetical generation set re-sampling,which improves the operation efficiency while also improving the better matching accuracy.The last step is the restoration of three-dimensional coordinates and camera parameters.The principle and steps of constructing sparse point cloud with incremental structure from motion(SFM)are studied.Aiming at the problem that the serial iterative reconstruction strategy is easy to reduce the reconstruction efficiency,the reconstruction strategy based on the three-view geometric constraint hierarchical iteration is proposed.Experiment shows that the hierarchical iterative reconstruction algorithm obtains more reconstructed point clouds while increasing the reconstruction speed.