| SLAM systems based on monocular vision have made some research progress,but there are problems that need to be solved,such as lack of scene information,overly sparse maps and difficulties in initializing monocular cameras.In order to overcome the above difficulties,this paper proposes a semantic map construction system based on object detection and monocular visual SLAM,and the main work is as follows.(1)The framework of traditional visual SLAM system is investigated,and a method of building semi-dense semantic maps with monocular cameras is proposed for the defects of SLAM system using only monocular sensors,which can use low-cost monocular sensors to build maps that can help mobile robots locate and understand the environment and get better practical applications.(2)YOLO v4 is used to detect and output the bounding box coordinates,confidence level and semantic class information of different objects.The 3D coordinates of objects are generated from 2D bounding boxes by the vanishing point algorithm and the quadratic surface recovery algorithm based on the object detection results to achieve tracking of objects and introduce semantic information into the maps constructed by SLAM.(3)In addition to object tracking,the outlier points are removed by integrating the isolated forest and local anomaly factor algorithms,which enables the system to complete the 3D object initialization more accurately and recover more accurate object poses;the length and direction invariance of the abstract line segments are introduced,and the constrained,BA-optimized loss function containing geometric and semantic relationships is constructed.The camera poses and 3D point positions are optimized by minimizing the minimum error between points and objects in the observed and predicted images to further optimize the camera poses.Finally,experiments are completed in the TUM public dataset and real scenes and compared with monocular ORB-SLAM2 and other monocular SLAM systems.The SLAM system proposed in this paper can accurately extract the semantic information of objects in the environment,and the semi-dense maps constructed can better express the environment than the sparse maps constructed by the monocular systems widely studied so far,with improved localization accuracy. |
