Research On Semantic Mapping Method For Large Scale Outdoor Environments

In recent years,robots have played a very active role in the execution of outdoor environmental tasks such as public safety and disaster relief,and the market size has gradually expanded.However,the ability of robots to perform tasks autonomously is very limited.Especially in the process of simultaneous localization and mapping(SLAM)of a dynamic environment,the robot lacks the perception of the semantic information of the environment.At the same time,it is interfered by dynamic objects and easily loses localization information,such that it has difficult to perform high-level intelligence tasks.Therefore,in order to improve the robot's ability to understand the environment,this paper studies the semantic mapping method in the dynamic environment from the following three aspects,including:(1)This paper proposes a method for semantic mapping of outdoor environment based on binocular stereo vision.This method obtains the disparity map using the ELAS algorithm and the DeepPruner network respectively,and uses the semantic segmentation network to obtain the pixel-level semantic information of the image.When using the disparity obtained by DeepPruner to fill the holes in the ELAS result,some objects with strong texture are replaced according to the results of semantic segmentation.The method in this paper can get a smoother and more beautiful result when building a semantic point cloud map.(2)In order to reduce the impact of the interference of dynamic objects on visual odometer in outdoor environment,this paper proposes a visual odometer method based on dynamic feature removal.Since the traditional ORB-SLAM algorithm shows an excellent performance in a static environment,this paper proposes a dynamic feature detection and removal module on the basis of ORB-SLAM,which improves the accuracy of ORB-SLAM's pose estimation in a dynamic environment.This module uses the YOLOv3 network to detect movable objects in the image as a candidate area,and combines the epipolar detection method to detect the dynamic feature points in the image,and finally uses the optical flow method to track the detected dynamic features.By demonstrating the experiments on the public data set of KITTI.this method has better performance when there are dynamic objects in the environment.(3)Due to the low efficiency of map building in large outdoor scenes and memory limitations,this paper proposes a divide-and-conquer merging idea to map the entire large-scale scene map.When mapping an outdoor scene map,the image sequence of the environment is synchronously collected at different locations to map a sub-map,and the similarity matching of the image sequence is detected based on the twin network,and the relative pose between the two sub-maps is calculated according to the candidate matching pair.Finally,the sub-maps are merged into the entire map.Using the same method,build a semantic point cloud map based on the results of the semantic segmentation network.In this paper,by introducing semantic information,combined with the principle of epipolar constraint to eliminate dynamic feature points,traditional SLAM can deal with relatively complex dynamic environment,and we propose a large-scale outdoor semantic mapping method based on the idea of divide and conquer merging to help robots understand the surrounding environment and execute high-intelligence tasks.