Three-dimensional Localization and Mapping of Multiagricultural Scenes Via Hierarchically-Coupled LiDAR-inertial Odometry
Yuanqian Hong,Ruijun Ma,Chenghui Li,Chengji Shao,Jian Huang,Yunyu Zeng,Yu Chen
DOI: https://doi.org/10.1016/j.compag.2024.109487
2024-01-01
Abstract:Accurate situational awareness is crucial for various agricultural tasks and serves as a foundation for autonomous agricultural robots. Simultaneous localization and mapping (SLAM) based on vision and Light Detection And Ranging (LiDAR) has been applied to agricultural robots, such as state estimation and environment perception; however, most of them adopt feature extraction in the front-end. Owing to disruption in feature extraction in agricultural environments and the dependence on edge and surface features in traditional LiDAR SLAM, system accuracy and stability can rapidly decline. Herein, we introduce a hierarchically-coupled LiDAR-inertial odometry system without feature extraction and a comprehensive SLAM system that incorporates a factor graph via loop closure detection. Initially, points originating from the robot and outliers are removed using crop and statistical filters. Subsequently, inertial measurement unit (IMU) measurements are used to diminish the pointcloud distortion caused by moving and turning. Third, we present an enhanced voxelized generalized iterative closest point method that employs an adaptive probability distribution for front-end pose estimation. Finally, the back-end is hierarchically incorporated with a factor graph and geometric observer to optimize the global states of the LiDAR key and common frame. This includes updating the bias of the IMU's state and the robot's speed state in preparation for next-frame de-distortion. We conducted experiments on four self-collected datasets of a dragon fruit field, a sugarcane field, a mango orchard, and a mixed scene with buildings, as well as two public datasets of a park scene and a street scene. In the mango orchard and mixed scene datasets, we observe derivations of 0.1357 and 0.0564 m, respectively, by returning to the starting point method. For the two public dataset sequences of a park scene and an outdoor street, the absolute positioning RMSEs of our method are 0.4371 and 0.1253 m, respectively. Compared with state-of-the-art benchmarks, our method offers more competitive performance in agricultural scenes. In addition, our approach is immediately adaptable to LiDAR sensors with different scanning models in the case of featureless design. These results demonstrate that our method exhibits good localization and mapping performance in open or dense agricultural environments. This exploratory research provides theoretical and technical references for SLAM in future universal agricultural robot platforms. It has broad basic application capabilities and can provide the necessary technical support required for smart agriculture.