Research On Mobile Robot's SLAM Based On Depth Camera

Simultaneous localization and mapping(SLAM)technology has a wide range of applications in the fields of autonomous driving,service robots,drones,AR/VR and SLAM is the key for agent to realize the autonomous navigation.The depth camera represented by Kinect has been widely concerned by SLAM researchers.This thesis,whose mobile robot vision SLAM research is based on the depth camera-based,has important theoretical significance and application value.The main tasks are as follows:Firstly,this thesis compares the advantages and disadvantages of laser and visual SLAM sensors,selects Kinect camera as the environmental information acquisition sensor of visual SLAM,studies the imaging principle of Kinect camera and the calibration method of camera internal reference,and completes separately internal reference calibration of the color,depth camera under Robot Operating System(ROS).The internal reference of the depth camera is calibrated.For the indoor environment,this thesis improves the RGB-D SLAM scheme based on the dominant modules in the common visual SLAM scheme.Secondly,in the front-end pose estimation,a mismatch matching method is proposed for the noise problem in the depth information acquired by the Kinect sensor.This method uses the Gaussian function as the loss function and uses the weight classification sample RNASAC algorithm to match the image accurately.Then,this thesis combines the depth information of the image matching point set and the threshold filtering to improves the inter-frame motion estimation by the Iterative Closest Point(ICP)algorithm registration.Experiments show that this method improves the image matching accuracy and reduces the average time consumption of the algorithm,and increases the motion transform estimation.Then in the back-end optimization,for the drift problem occurs in the motion trajectory between adjacent frames,the thesis adopts the Bundle Adjustment(BA)algorithm to partly optimize the pose estimation between adjacent key frames.For the effect of closed-loop detection,the thesis combines the improved Bag-of-Words(BoW)spatial segmentation and the ICP-registered interior point threshold to filter loopback frame.Based on the closed-loop detection,the Ceres method is used to optimize the global estimation of pose estimation.The experimental results show that the BAalgorithm reduces the motion trajectory drift.On the basis of the improved closed-loop detection,the pose pattern optimized by the Ceres method is closer to the real data,which improves the accuracy of the motion trajectory estimation.Finally,the realization of Kinect-based mobile robot SLAM system is completed on the hardware platform of Pioneer-3DX mobile robot and the software platform of ROS operating system.In order to verify the improved pose estimation method proposed in this thesis,improve the accuracy of the closed-loop detection method and the feasibility in the indoor environment,experiments are carried out in the TUM data set and indoor scene respectively.The experimental results show that the proposed method reduces the pose estimation error of mobile robots,and it is feasible to locate and construct the indoor environment.