The Research Of Positioning And 3D Construction Method Based On Heterogeneous Platform

In recent years,with the rise of drones,unmanned vehicles,service robots,storage robots,and AR,localization and 3D reconstruction as one of the core technologies has become increasingly important.Laser-based and GPS-based sensors are difficult to use because of their high cost or other limitations.Simultaneous positioning and dense construction based on vision sensors have become the mainstream.With the help of binocular cameras for dense depth recovery,the parallax of the entire picture needs to be calculated.The time cost of calculation is very large,and therefore it is not possible to build dense map in real time.At the same time,the estimated pose error is larger due to the matching error of the feature point method.This thesis includes restoring image depth and solving motion trajectories.In the depth recovery part,the GPU is used to accelerate the parallax algorithm to obtain the depth information.In the trajectory and positioning part,after analyzing and comparing the optical flow method and the feature point method,the feature point method based on the ORB feature is used because the optical flow method could not be used outdoors,and use PnPRANSAC method to get the trajectory.In order to improve the stability of the system,this paper designs a local feature map to implement feature matching of feature points.Then the overall optimization is performed using the graph optimization framework based on the least squares method.Finally,through the depth information obtained by the GPU acceleration and the pose information obtained by the tracking module,dense space construction is achieved by means of the PCL point cloud library.According to experiments,the GPU-accelerated binocular disparity algorithm can process pictures in real time to obtain the disparity map,and the improved feature point method matching accuracy has been greatly improved.The binocular data set is used as the real data to simulate and test the algorithm.The algorithm realizes the function of simultaneously map and positioning and obtains 3D map and camera trajectory.After comparing the true trajectory with the trajectory estimated by this paper,it is clear that the camera trajectory is calculated correctly and the error is stable within a very small range.After accelerating the depth recovery module,the speed of the algorithm based on heterogeneous platforms has been greatly improved.