Semantic map generation algorithm combined with YOLOv5

Abstract

Traditional visual SLAM algorithms have problems such as lack of semantic information, low accuracy and slow speed of 3D point cloud segmentation. This paper proposes a semantic map generation algorithm based on YOLOv5 and improved VCCS point cloud segmentation. Firstly, the ORB-SLAM2 algorithm is used to generate the original three-dimensional point cloud. The target is detected by YOLOv5 and the original point cloud is semantically annotated, and the objects in the point cloud are expressed in other colors. Then, the VCCS algorithm was used for over-segmentation to obtain supervoxel clustering. The improved VCCS algorithm was used to merge supervoxel clustering to improve the accuracy of segmentation results. Finally, a three-dimensional point cloud map with semantic information is established. Experiments show that the algorithm can generate semantic maps very well, and the accuracy and speed of 3D point cloud segmentation are greatly improved.