Research On 3D SLAM Of Mobile Robot With Omnidirectional Stereo Vision

Recently, omnidirectional imaging system has become a hot issue of the research on robots’ visual field, this camera has been widely used in visual inspection field, robots’ location and servo control system of robots’ vision. With the ability of robots has been improved and indoor environment has become more complex, in order to decide robots’ moving strategy, 3D information of the environment should been taken into account. This paper studies 3D SLAM based on omnidirectional camera, the purpose of is to deeply solve core principle of 3D SLAM, prove the validity of research method on the robot platform, and offer technology support for high level environmental cognition of robot.Before using the camera, we should calibrate it firstly. A calibration method is proposed for fish-eye omnidirectional camera. We adopt the unifying spherical projection model to represent the optical nature of our camera, and build binocular stereo vision projection model. Traditionally, Harris and FAST corner detection work unreliably in highly distorted image caused by viewport variance. Therefore, edge-based corner detection is proposed based on Hough transform. The line segments within projected window and their intersection corner points are extracted accurately and reliably due to the robustness of Hough transform.The achievements of the research contain 3D reconstruction of SIFT feature points and the 3D SLAM algorithm. Because the traditional SIFT algorithm is not suitable for omnidirectional camera, a SIFT algorithm in spherical coordinates for omnidirectional images is proposed in this paper. This algorithm can generate Local Spherical Descriptors, with the descriptor, point matching between omnidirectional images can be performed. And also studies modeling and detect method of 3D planar feature, and the method of obtaining the 3D information of spherical SIFT feature points. Use RANSAC algorithm to plane fitting the obtained 3D points cloud, and prove correctness of this method. Solve the SLAM problem of the stereo vision system based on Extended Kalman filtering. In contrast to most existing approaches to visual SLAM, the present method does not rely on restrictive smooth camera motion models, but on computing incremental 6-Do F pose differences from the image flow through a probabilistic visual odometry method. Moreover, our observation model contain both 3D positions and the SIFT descriptors of the landmark, and use Kd tree algorithm to solve the data association problem which is an important part of SLAM. Two experiments are shown in this paper to test the performance of the proposed SLAM method, and achieve 3D SLAM on a robot platform.The results of experiments show: spherical SIFT algorithm successfully overcome the problem that traditional SIFT algorithm behave badly on spherical pole extraction, it is suitable for omnidirectional camera, and improve the matching effect of omnidirectional images. Use planar fitting method to prove the validity of obtaining 3D feature point information method. The proposed 3D SLAM method is suitable for different environments, and the error of constructed maps is low, and could be transplanted to robot platform.