Slam Based Quasi Dense Three-Dimentional Reconstruction

The goal of visual system based three-dimensional reconstruction is to extract the depth information of scene from the sequence of images or video.Monocular camera is the most widely used in visual system,but it lost depth information during imaging.In recent years,the three-dimensional reconstruction based on the Simultaneous Localization and Mapping(SLAM)system has drawn the attention of scholars.However,the existing systems have many problems,such as excessive demands on the hardware,inability to meet the real-time requirements,and sparse reconstruction models.Therefore,in this paper,we studies the disadvantages of the existing SLAM system to ensure the system efficiency as much as possible while increasing the density of the model so as to meet the human visual sensory needs.Papers mainly include the following work:First of all,in order to solve the problem that the original ORB-SLAM system can only reconstruct the three-dimensional information of the feature points,the direct method and the indirect method are combined to obtain the pose and pose information of the camera.Then the relationship between the key frames is obtained through the pose information.The matching relation is found between the adjacent key frames when the gradient is greater than the threshold value,and the obvious region is reconstructed using the probability model.In addition,the system also adds depth information module to get more depth information,so that the recognition of the results are greatly improved.Secondly,in the aspect of algorithm optimization,the traditional method of polar line search uses the method of gray value comparison.This method has poor matching in the case that the difference of two image brightness values is great.In this paper,the gray and gradient are taken as the matching requirement,which can greatly improve the matching accuracy.And new keyframe selection criterion is introduced to limit the pose and position relationship between the neighbor frame.Finally,in order to improve reliability,we propose the improved bilateral filter to enhance the depth image estimated by monocular camera.And then the intra-frame and inter-frame error rejects are introduced to cull the error points,which improves the accuracy of the improved system reconstruction remarkably.