| With the increasing demand for interaction between robots and the environment,there is a growing tendency to explore the environment using visual SLAM(Simultaneous Localization and Mapping).However,most visual SLAM algorithms have problems such as point cloud being too sparse,lacking semantic information,and low practicality of the constructed map.In order to solve the above problems,the thesis proposed a semantic map construction system based on visual SLAM and object detection,the main research contents include:(1)In terms of semantic map construction,we proposed a method which combine the object detection algorithm YOLO v4 with visual SLAM to construct a dense point cloud semantic map of the environment.According to the proposed semantic SLAM system,YOLO v4 is used to obtain the 2D labels of the object,and the sparse point cloud map of the environment is constructed through the ORB-SLAM2 algorithm.The system combine 2D labels of the color image,corresponding depth map and the key frame to generate the dense point cloud label.Then,we use a graph-based segmentation algorithm to segment the dense point cloud,and then merge the point cloud labels with the segmented point cloud to construct a dense point cloud semantic map.(2)In terms of obtaining object semantic labels,the trained YOLO v3 and untrained YOLO v4 algorithms are analyzed from the three aspects: detection accuracy,time consumed and recognized categories.We experimented the two algorithms in the actual environment to verify the necessity of choosing YOLO v4.(3)In order to improve the visualization of the point cloud map,when performing texture mapping,we select several multi-view RGB images to texture the geometric model to reduce the seam texture between adjacent texture blocks to ensure the consistency of the texture.(4)In order to reduce the storage space of the point cloud map and facilitate mobile robot to avoid obstacles and navigate,we use semantic octomap as another representation method of semantic point cloud maps.Finally,we use the TUM public data set and indoor actual scenes to test the proposed method.Compared with the traditional ORB-SLAM2,the proposed system reduces the absolute pose error and the absolute trajectory error of the camera in the process of constructing the map,which improves the accuracy of mapping.The results of our experiments show that a practical and highly visible semantic map can be constructed by our system. |
