Research On SLAM Algorithm Of Mobile Robot Based On Depth Camera

Simultaneous Localization and Mapping of mobile robots is a key technology to intelligent robot.When a robot is in an unknown environment and unable to get its location information from outside,it needs to build a map of the environment and locate to it based on mounted sensors.Robot localization mainly relies on sensors such as lidar,IMU,distance sensor.With the development of computer vision,cameras are widely used in the field of SLAM because of its low cost,low power and data richness.It is called visual SLAM.Microsoft's RGB-D camera,launched in 2010,has created a new research boom in visual SLAM.In this thesis,we improved the RGB-D SLAM published in 2012 with the combination of some latest SLAM algorithm and modules.We also use RGB-D camera as the only sensor and the algorithm is mainly used to test indoor environment with rich feature points.This paper analyzes the merits and demerits of direct method and feature-based method in camera pose estimation,improves the camera tracking module of the front end,we propose a coarse-to-fine pose estimation method.In the RGB graph,ORB features are extracted.First,the sparse direct method is used to track these ORB features to get the rough estimation of the current frame.A certain key frame selection mechanism is set up.The candidate key frames are sent to the position estimation part,and the feature point method is used to estimate the position and posture and to select the key frame.Aiming at the camera tracking module,this paper improves the loop detection method and key frame strategy in RGB-D SLAM.We use a graph optimization method in the back end to optimize the global pose of key frames.And transform the three dimensional point cloud of key frames to the world coordinate system to achieve the final three-dimensional dense point cloud map of the indoor environment.The algorithm is tested on the open dataset of TUM.The experiment proves that this algorithm has faster computing speed than RGBD SLAM while ensuring high position estimation precision,and the algorithm can realize real-time operation on CPU.Finally,pose estimation and 3D dense point cloud map reconstruction can be realized.