Study On SLAM Technology Based On Stereo Vision

SLAM(Simultaneous Localization and Mapping)has always been a key technology for robotic navigation and positioning.Compared to monocular and RGB-D cameras,the use scene of the binocular camera is not limited,and the depth value of the pixel can be calculated,so that the constructed map has scale information.The direct method based SLAM technology has made great breakthroughs in recent years.Therefore,this thesis mainly studies the binocular vision SLAM method based on direct method.The specific research content of this thesis is as follows:(1)The front end of the visual SLAM system based on direct method is studied.Firstly,the method of photometric calibration can solve the problem that the gray-scale invariant assumption is not established in the actual scene.Secondly,the tracking method based on direct method is studied.The pyramid region segmentation method in feature point method is applied to the pixel extraction process of direct method to enhance the robustness of system tracking.The key frame mechanism is applied to reduce the computational complexity of the system.Combining a constant speed motion model with a reference key frame model provides a good initial value for backend optimization.Finally,depth estimation is performed by combining multi-view geometric depth estimation method and binocular depth estimation method.(2)The back-end optimization of the SLAM system was studied.Firstly,a photometric error model based on direct method is constructed,and the sliding window optimization method is used to optimize the pose and map points.Secondly,the Shi-Tomasi corner detection process was added.The closed-loop detection based on the word bag model is merged into the direct method system.Finally,the global optimization model is studied to optimize the global pose.(3)The SLAM system is built.The effectiveness of the pyramid region segmentation method,the depth estimation method and the closed-loop detection of the fusion is verified by the data set simulation experiment.Then the system built in this thesis is tested and verified in the real scene.The system can locate and build semi-dense maps in real time.