Research On Dense 3D Scene Reconstruction Based On Monocular Image Sequences

Reconstructing the three-dimensional(3D) structure from multiple image views has become an important research topic in the domain of computer vision. The target objects 3D information come from 2D image sequence and cameras' parameters. Compared with the traditional methods such as the 3D scanner or the modeling software, 3D reconstruction algorithms based on image sequence have advantages on fewer constraints, higher degree of automation, lower cost and meeting requirements for large-scale scene reconstruction. This technique has a wide application prospect in many fields like reverse engineering, mobile robot navigation, virtual reality, augmented reality, augmented virtual environment, military affairs and so on.Nevertheless, long-term practice indicates that the 3D reconstruction algorithms based on image sequence are confronted with many problems, such as camera external parameters calibration technology with higher complexity and lower accuracy is insufficient to provide an accurate estimation of the camera pose, inaccurate feature matching results cannot provide accurate three-dimensional information.Under the non-cooperative environment, the image feature detection and matching algorithms are difficult to detect reliable matching feature points in the weak texture area or texture repeated region, which causes larger distortion in resulting point cloud and surface reconstruction levels. The reconstruction surface appears serious hole, which became prominent technical barriers.In addition, the feature matching process on the number of feature point and the speed of operation is difficult to take into account. This paper studied the above issues, the main contributions are the following.(1) Designed a camera pose tracking algorithm based on structure from motion method, using SIFT algorithm to detect and match image feature points, using five-point algorithm based on RANSAC framework to estimate the two-view relative pose, using Perspective-n-Point algorithm based on RANSAC framework for multiple view pose tracking, these effectively improve the pose estimation accuracy.(2) A monocular dense 3D reconstruction technique based on optical flow feedback is proposed to achieve accurate and rapid 3D stereoscopic modeling in the real scene. Corresponding pixel pairs are robustly matched by TV-L1 optical flow algorithm, the dense matching points are sampled and then sparse point cloud is generated and initial coarse mesh is built. In the proposed method, multi-view reconstruction is implemented from perspective of vision method on motion analysis. The reconstruction model is fed back to the reconstruction process, and the model is deformed by utilizing the bias-driven of each view. The coarse and inaccurate original mesh surface is adjusted to the exact surface through a dense non-rigid deformation.(3) Under the Compute Unified Device Architecture(CUDA), the optical flow algorithm, synthesized image codes and mesh adjustment algorithm are optimized in parallel mode by using the Graphic Processing Unit(GPU) hardware, and real-time performance of the reconstruction algorithm is significantly improved.In this paper, an effective 3D reconstruction solution is put forward for improving the accuracy of camera external parameters, feature matching accuracy, the denseness of 3D reconstruction result and reducing the time consumption. The experimental results obtained in realistic indoor scenario demonstrate the effectiveness and accuracy of the proposed algorithm.