Indoor 3D Scene Reconstruction Research Based On Dense Vision

The 3D reconstruction for indoor scenes originated from the problem of Simultaneous Localization and Mapping(SLAM).It is a technology based on sensor to collect real scene information and reconstruct it on computer.It is widely applied in many fields,such as augmented reality,virtual reality,artificial intelligence and so on.In recent years,with the development of SLAM technology,the three-dimensional reconstruction of indoor scenes has also achieved great achievements.However,due to the large amount of data and complex environment,the algorithm still has some problems in operation efficiency,accuracy and reconstruction effect,and there is much room for improvement.Based on this,we propose an indoor scene reconstruction algorithm based on dense vision,which can effectively reconstruct the real indoor environment.The main contributions of our research are as follows:1)In order to solve the problem of the camera motion's computation.We design a dense method using all pixels to estimate the motion of the camera.Compared with the method based on feature point which only uses feature information,our algorithm runs faster and has higher precision.The algorithm is similar to the optical flow method in principle,but at the computing stage,we use a probability method to minimize the error.Compared with general optical flow algorithm,our research not only reduces the dependence on the assumption of constant flow,but also can calculate the larger camera movement through a coarse to fine way.2)In order to solve the problems of tracking loss and error accumulation in the process of computing the global motion trajectory of a camera.We design a dense key-frame selection algorithm based on random fern cluster classification.First,we classify and code images by the fern cluster.Second,we get the key-frame sequence by comparing the difference between frame coding,and recover the missing image according to the sequence.Finally,we construct an optimization model based on the key-frame sequence obtained before to reduce the cumulative error.Compared with the general key-frame algorithm,our algorithm has a better effect on the correct rate of key-frame lookup and the tracking loss rate.At the same time.Compared with the general optimization model,the optimization algorithm based on key frame is more efficient.3)In order to solve the problem of the three-dimensional reconstruction for the indoor scene in the computer.We propose a modeling method based on octree.By using the octree structure,we save the 3D information of the scene,fuse data and complete the three-dimensional rendering.Compared with the traditional method based on point cloud or voxel,our method has a better comprehensive performance and is more suitable for modeling the indoor scene.Through theoretical analysis and experimental verification,our research better solves the accuracy and efficiency problems of indoor reconstruction,also make contributions to the development of 3D reconstruction technology.