Research On The Algorithm About Fast 3D-Point Cloud Reconstruction

3D-Point Cloud reconstruction is to restore the three dimensional structure of physical scene or object model by using hardware(such as camera,laser etc.)to get depth information,directly or indirectly.In this paper,we propose a real-time 3D-Point Cloud reconstruction method by using monocular camera.The involved camera tracking problem is solved by using monocular based simultaneous localization and mapping algorithm(monoSLAM)The monoSLAM algorithm proposed in this paper includes initialization,feature based tracking,relocation,local optimization,update and propagation of depth and loop closure.At first,we use an initial optimization strategy,which can fix the relative scale and this guarantee for the depth update and propagation running smoothly.In local optimization module,we use a set of optimization strategies to improve the accuracy of camera pose tracking and make the real-time depth updating possible.In order to reduce the computational cost of depth updating,we also apply the depth updating strategy to new pixels and undistort every frame at the very beginning.In addition,the ways of data interaction among multi-threads in our system are also explained in detail.In order to make a compare with depth update and propagation method,we implement the semi-global matching algorithm(SGM)which have been speeded up and source code has been submitted to github.At last,the accuracy of tracking trajectory and performance of reconstruction are evaluated in TUM RGB-D benchmark and our own datasets,which proves our system reconstructs semi-dense point clouds of the 3D scenes with higher accuracy in real-time.