AELVI-SLAM: LiDAR–visual–inertial SLAM for autonomous exploration in multiple agricultural scenarios

Accurate simultaneous localization and mapping (SLAM) algorithms are essential for enhancing the operational efficiency of agricultural robots. However, existing SLAM methods are vulnerable to dynamic disturbances and cumulative drift in complex agricultural environments. Moreover, traditional SLAM pipelines typically focus on state estimation and mapping, while map construction and data collection often rely on manual or pre-planned exploration, which is time-consuming and insufficiently automated. Therefore, this study proposes a LiDAR–visual–inertial SLAM framework (AELVI-SLAM) to enable autonomous exploration and mapping for agricultural robots in unknown environments. An object-level dynamic removal strategy is introduced. Building on clustering, object segmentation combines reflectance intensity and geometric features, and dynamic targets are removed using a consistency metric with weighted residuals. A deep learning-based loop closure detection method is introduced within a factor graph framework that fuses IMU preintegration, LiDAR-visual odometry, and loop-closure factors to improve long-term trajectory consistency and mapping accuracy. Finally, the autonomous exploration algorithm is integrated and deployed on an agricultural robotic platform, and field tests are conducted in representative scenarios including greenhouses, farmland, and orchards. The results show that AELVI-SLAM effectively filters dynamic objects and outperforms state-of-the-art methods in both mapping quality and trajectory accuracy. Specifically, the standard deviation (STD) of the absolute trajectory error (ATE) is less than 0.0205 m, the root mean square error (RMSE) is less than 0.0362 m, and the average processing time is 209.82 ms, satisfying real-time requirements. The integrated autonomous exploration algorithm enables path planning across three agricultural scenarios. The constructed colorized point cloud maps enrich environmental perception and support crop phenotyping, semantic mapping, and the planning of agricultural field operations. • A LiDAR–visual–inertial SLAM framework was proposed for autonomous exploration in agricultural environments. • An object-level dynamic removal method improved robustness against dynamic disturbances. • A deep learning–based loop closure integrated with factor graph optimization reduced long-term drift. • Field experiments demonstrated improved mapping quality and localization accuracy in real agricultural scenarios.