Vehicle 3D Modeling Technology Based On Invisible Structure Light

3D modeling technology based on structured light is a research and application hotspot in the computer field due to its high reconstruction efficiency and its ability to be used for objects lacking texture features.However,its projected pattern is easily disturbed by light in the scene and the modeling size is limited.A project is designed by exploiting infrared structured light and point cloud stitching in this thesis,aiming at solving these issues and implementing 3D modeling of physical vehicles in outdoor scenes.The project includes three steps: collection of vehicle point cloud,simplification and triangulation.(1)Vehicle 3D point cloud collection.We designed and manufactured a rigid gantry with infrared depth sensors to collect point cloud data from multiple acquisition angles.According to the rotation and translation transformation in the space rectangular coordinate system,the point cloud coordinates are normalized to obtain the point cloud of the entire vehicle.(2)Vehicle point cloud sampling and denoising.Firstly,a curvature-based sampling algorithm is proposed,which constructs triangles with K-means clustering results in each grid,and uses the mean value of the adjacent triangle normal angles as the local curvature estimation value to retain more 3D points in the high curvature regions of the vehicle.Compared with uniform sampling,this method achieves more accurate performances on describing the contours of the vehicle.For the limitation of erroneously sampling by the aforementioned artificially sampling rules,a deep convolutional neural network for vehicle point cloud sampling is proposed.Point clouds in each grid are fed into the deep convolutional neural network for channel scaling and feature mapping,to exploit the structural commonality of vehicle point clouds.After the training phase,the feedforward neural network is then utilized for point cloud sampling,which enhances the nonlinear representation ability of the point cloud sampling process.Finally,according to the spatial position,the dense noises on the surface are filtered out,and the cubic data statistics are performed along the coordinate axis to filter out the sparse noises around the vehicle.(3)Vehicle point cloud triangle meshing.According to the Delaunay triangulation,the vertex constraint and assessment function are formulated in this thesis.Starting from an initial triangle,new triangles are continuously generated by selecting point of side with largest evaluation value,thereby the triangular meshing of vehicle point clouds is incrementally completed,and finally a closed surface is generated for building the 3D model of the physical vehicle.Experiments show that the 3D modeling scheme designed in this thesis effectively solves the shortcomings of structural optical coding information being easily interfered by the weakened lights and limited by the modeling size.In outdoor environment,the vehicle model can be built in less than one hour,and compared with the standard size,the average error of length,width and height is less than 1%.