Research On Semantic SLAM Algorithm Based On Movable Object Removal

Simultaneous Localization And Mapping(SLAM)refers to the use of some devices such as cameras or radars to obtain external information to determine the pose of the robot in the environment,while establishing a map of the explored area.Most of the existing slam solutions are carried out in the environment of constant illumination,and the camera is used to obtain rich color data.In the outdoor environment,the mapping and positioning effect is not good.Moreover,there are a large number of moving objects in the outdoor scene.Most of the features extracted by the feature extraction algorithm fall on the dynamic objects,resulting in positioning deviation and poor robustness of the algorithm.Therefore,it is necessary to deal with the dynamic objects in the real environment.In this paper,a dynamic object identification algorithm is designed for the outdoor real road environment,which processes the moving objects,and introduces semantics as the auxiliary information of positioning,which can ensure the positioning accuracy and improve the real-time performance of the algorithm,making the algorithm more robust in the actual environment.Firstly,according to the characteristics of feature extraction in slam,this paper analyzes the influence of mobile objects on odometer.Based on the existing three-dimensional point cloud semantic segmentation algorithm,using the speed of the object and a priori moving probability,a moving object identification algorithm is designed.By iterating the moving probability of the object,the moving object is removed,and its influence in slam pose estimation and mapping is reduced.Secondly,the artificial features are easily affected by the observation location,weather and other factors,and the observation features are reduced after removing the movable objects.Therefore,this topic adds the semantics of point cloud and the global features of point cloud as the auxiliary location information.Semantics is a high-level feature of human understanding of the world,which will not be affected by the above factors.In this project,we add the category label of the object on the basis of the artificially designed feature vector,and add three global features of the point cloud: the total number of objects in each frame,the nearest distance from the lidar,and the farthest distance from the lidar.The new features make the feature vector more robust to ensure the accuracy of feature matching when the features are reduced.Finally,we build a complete semantic slam system for moving object removal,and do some quantitative experiments on the timeliness and accuracy of SLAM Algorithm in Kitti dataset,as well as qualitative experiments on the mapping effect.The experimental results show that the semantic SLAM algorithm designed in this paper can improve the real-time performance by 18.78% and 16.0% on average when the positioning accuracy reaches the same level as segmap.Compared with squeezeseg's 83.2%,the recognition accuracy of the proposed algorithm is 91.7%.The semantic information is introduced to improve the robustness of localization.The algorithm proposed in this paper can run well in outdoor dynamic scene and build environment map without dynamic objects.