Research On Semantic SLAM Based On Fusion Of Vision And Inertial Information

In recent years,with the continuous progress of science and technology,robots gradually develop towards intelligence.The traditional system of SLAM has been unable to meet the high-level task requirements.Robots need to complete tasks from the semantic level,so how to integrate semantic information into the SLAM system is the key to improve robot capabilities.In this paper,the application of environment semantic information in SLAM system is systematically studied for outdoor scenes.The main contents of this paper are as follows:Aiming at the problem that the traditional monocular SLAM system ignores dynamic objects in the scene,this paper fuses semantic information into the SLAM system and removes dynamic feature points through semantic information.First,in terms of semantic segmentation,this paper proposes a CRF image semantic segmentation algorithm based on FCN and superpixel.This algorithm optimizes the semantic coarse segmentation map through the effective edge information of superpixel,and refines the segmentation edges through CRF.Based on the results of image semantic segmentation,this paper proposes a dynamic feature point detection algorithm based on semantic information and polar constraint.The algorithm obtains the potential moving object through the result of semantic segmentation,eliminates the dynamic feature points on the object through the polar constraint relationship,and finally uses the static feature point to estimate the pose,improving the robustness of SLAM system in dynamic environment.The experimental results show that the proposed algorithm is more accurate than the mainstream algorithm in dynamic scene.Aiming at the problems of traditional monocular SLAM system,such as losing the absolute scale information,and losing the target in the fast motion scene easily.This paper fuses IMU information into the monocular SLAM system,proposes an IMU-assisted optical flow tracking algorithm.In this algorithm,the relative motion between two frames is obtained by IMU pre-integration,and the relative motion is used as the motion constraint of optical flow to accelerate the iteration speed of optical flow method,so as to improve the operating efficiency of the system.On the basis of the visual inertia odometer,an IMU-assisted algorithm was proposed to eliminate the false matching points between two frames when the camera was moving fastly.Based on the results of IMU pre-integration,this algorithm calculated the Fundamental matrix,and calculated the distance between feature points and polar lines according to the Fundamental matrix,and then eliminated the points with large errors.Finally,it used the RANSAC algorithm to eliminate the mismatching points,effectively improving the problem of mismatching.The experimental results show that the proposed algorithm is more accurate than the mainstream algorithm in the scene of fast motion.Aiming at the problem that 3D dense maps constructed by traditional SLAM systems only contain low-level information such as color,depth,and brightness,this paper constructs the environment semantic map based on the traditional 3D dense map,associates the 2D semantic labels in multiple frames through Bayesian update,and transfers them to the 3D point cloud to get the preliminary 3D semantic map.On this basis,a global optimization algorithm based on higher-order CRF is proposed.In this algorithm,the high-order term of CRF is established by the 3D superpixel with the same time and space,and the constraint relationship between the point cloud and the 3D region is added to achieve the boundary consistency of the category of the point cloud in the semantic segmentation,so as to alleviate the influence of the over smoothing caused by the binary term,so as to improve the accuracy of the segmentation.Experimental results show that the algorithm can get globally consistent semantic map,and can effectively improve the semantic segmentation results of single image.