Research On Indoor Positioning Algorithm Based On Depth Camera

Visual simultaneous localization and mapping(SLAM)technology uses visual sensors for pose estimation and environment mapping,which is often used in mobile robots,virtual reality and automatic driving.The depth(RGB-D)camera in vision sensor can meet the accuracy requirements of visual SLAM because of its minimum measurement error,so visual SLAM based on RGB-D camera has gradually become a research highlight.In order to solve the problems of low real-time and poor accuracy of visual SLAM technology based on RGB-D camera in indoor environment,this paper studies and improves some key modules of traditional visual SLAM technology.Under the framework of ORB-SLAM2,the following works are carried out:1)For the problem that the feature points extracted by ORB algorithm are dense and have a great influence on the accuracy of subsequent feature point matching and camera pose estimation,this paper proposes an improved uniform distribution algorithm of ORB feature points.Based on the image pyramid obtained by adaptive partition,this algorithm designs an improved adaptive threshold FAST corner detection algorithm to detect the corners in the image,and uses the improved quad_tree to manage the detected corners,so as to improve the computational efficiency of the uniform distribution algorithm.Finally,the experimental results indicate that the feature points extracted by the proposed algorithm are more sparse and globally representative than those extracted by ORB-SLAM2.2)For the problems of unstable iteration times and low real-time of RANSAC algorithm in traditional visual SLAM technology,this paper proposes a novel PROSAC algorithm.Compared with RANSAC algorithm,it randomly samples all the matching point pairs.The algorithm samples from the increasing pairs of correct matching points and calculates the homography matrix.According to the solution of the matrix,the wrong matching is eliminated,which largely raises the computational efficiency of the elimination algorithm.Experimental results show that the improved PROSAC algorithm is more efficient than RANSAC algorithm,and the effect of eliminating mismatches is better.3)For the problem that the key frames selected in the current visual SLAM technology can not reflect the global environment map information better,this paper proposes a threshold candidate key frame selection algorithm based on the image similarity.Firstly,all the image frames whose interior parameters are lower than the set threshold when matching with the feature points of the previous key frame are set as candidate key frames,and then the qualified candidate key frames are selected as key frames by using the rotation angle and displacement length,so that the key frames selected in this way can better reflect the global environment map information.4)There are many similar objects in indoor environment,which makes it easy for the front-end to recognize the similar position as the same position,that is,the front-end passes a false positive closed-loop constraint to the back-end,resulting in the failure of camera trajectory estimation.To solve this problem,this paper proposes a robust back-end optimization algorithm based on graph optimization to identify and eliminate the false positive closed-loop constraint,then the robustness of the loop closing detection algorithm is improved.Moreover,aiming at the problem of camera tracking failure,a relocation algorithm based on visual bag of words model is proposed,which transforms the relocation into image contrast problem.5)Based on the open indoor environment data set,simulations are carried out to verify the effectiveness of the proposed algorithm.Compared with the results of ORB-SLAM2 algorithm,the proposed algorithm is better than ORB-SLAM2 algorithm.