At present, a 3D reconstruction system with simultaneous localization and mapping (SLAM) based on the feature point method presents critical difficulties when the texture is missing. In contrast, with the SLAM based on the direct method, unsatisfactory reconstruction results are achieved when the camera moves at a high speed due to the difficulty in pose estimation. In order to solve such problems, this paper presents a dense 3D scene reconstruction system with a depth camera (RGB-D camera) based on semi-direct SLAM. The system uses the feature point method to estimate the pose of the camera in the rich region of the texture, and then uses an efficient incremental bundle adjustment to optimize the pose of the camera. In areas where the texture is missing, the direct method is used to estimate the pose of the camera. Therefore, the photometric error can be reduced when optimizing the pose of the camera. Then, a 3D map is constructed using the optimized camera pose. The surfel model and the deformation map are used to estimate the pose of the point cloud and the fusion point cloud. The 3D reconstruction system presents the following characteristics: (1) A hand-held camera can be used to scan for 3D reconstruction with any gesture, where the system can reduce the error of reconstruction model caused by human operation; (2) high robustness, stability and strength to deal with the jitter of the missing area and the camera; (3) dense reconstruction of a large scene can be performed, and the reconstruction effect can be well obtained. Multiple experiments show that the proposed system can be applied to 3D reconstruction of various typologies, and can get the optimal 3D reconstruction model. The obtained results prove its applicability in robot navigation, virtual reality shopping malls and other fields.

In order to obtain high-precision 3D models for 3D reconstruction of large low-texture scenes, a high-precision camera pose-estimation and optimization method is proposed in this paper. The method mainly uses a grid motion statistics feature-matching algorithm which calculates the number of matching points in the neighborhood to determine whether a match is correct or not. Therefore, this method can ensure that the initial value of the estimated camera pose has high accuracy. In the subsequent camera pose optimization, the image sequence is divided into several sub-sequences and each sub-sequence is independently assigned and optimized, which better solves the problem of camera pose drift verification when the cumulative error is gradually increased. The pose estimation and optimization method can not only obtain the initial value of the camera pose with high precision in the sparse region of the texture, but also solve the problem of reducing the cumulative error with the increase of the scene to obtain the high-precision camera pose when reconstructing the large scene. In our experiments, the size of the selected scene is generally larger than 100 square meters. The proposed methods and current state-of-the-art algorithms were compared quantitatively and qualitatively with published datasets and our own data sets in experiments. In six datasets, the average absolute trajectory error of the method in this paper is 0.014 m, which is smaller than Elasticfusion's result of 0.02 m (Elasticfusion is the method with the smallest pose error in the methods compared in this paper). The proposed scheme can obtain a high-precision camera pose and high-quality 3D reconstruction model, which can be widely applied in robotics, driverless vehicles and virtual reality.

No abstract

No abstract