Semantic Visual-Inertial Mapping System for Tree Maturity Determination

Abstract

We present a SLAM system, including a post-processing pipeline intended for tree maturity monitoring applications that uses stereo cameras, an IMU, and object information as pipeline inputs. The proposed SLAM methods based on ORB-SLAM2, produces a point cloud and localizes tree landmarks which are passed through the post-processing component to estimate tree trunk perimeters. We align landmarks with the world gravity vector, making it easy to extract tree trunk point clouds from the SLAM map in place of a direct 3D geometrical solution, which we find that would be computationally expensive. In experimental results, the SLAM system has good performance in SLAM pose estimation on the ETH3D dataset while also detecting object landmarks. In addition, the system can approximate tree perimeters, with a higher tree detection rate than a direct point cloud geometrical analysis approach.