Environmental Features Extraction And Scene Reconstruction Based On RGB-D Data

3D reconstruction can digitally reproduce real scenes of the objective world, which is an important research subject in the field of computer vision and robotics. As the development of application requirements,3D reconstruction methods are facing new challenges in real-time performance and reliability. This paper focuses on environmental features extraction based on the RGB-D data, develops research for problems of simplification for feature points, effectiveness evaluation for matching pairs, detection and optimization for loop-closure. The goal of this paper is realizing online3D scene reconstruction for high precision.For RGB-D data, this paper implements comparisons between a variety of image features and finally chooses ORB feature for the detection of scenes matching pairs, which performs good stability and real-time performance. In addition, considering the distance measurement error of sensors that results in the influence on the location accuracy of feature points, this paper proposes a depth uncertainty model, which can filter out feature points owning high uncertainty by calculating the depth mean and variance for each feature points. The depth uncertainty model reduces the number of matching pairs effectively and improves the credibility for feature points.This paper also solves the rotation and translation matrix between two scenes based on SVD algorithm. We test validity of feature matching pairs by using similar RANSAC algorithm. The scene matching degrees is calculated by matching degrees of multiple matrix sets of matching pairs in3D space. The highest value is chosen as the scene matching results.For the process of3D scene reconstruction, this paper proposes a global feature model for avoiding overlap, using Mahalanobis distance as association criterion and realizing model updating by uniting Kalman Filter framework. Because of accumulative error, we choose representative key-frames to detect loop-closure, which can be more efficient. Optimization algorithm based on graph is adopted to realize global optimization. Experimental results and data analysis verifies the validity and practicability of proposed method.