Adaptive navigation for robots in unstructured agricultural environments using stable feature localization and multi-sensor obstacle detection

Abstract

In this study, we present a robust and adaptive autonomous navigation system designed for unstructured agricultural environments, addressing challenges posed by localization instability and ground-level obstacles. The system consists of three main components: a prior point cloud mapping algorithm for environmental representation, a stable feature segmentation and matching algorithm for pose estimation, and a visual-inertial-LiDAR path planning algorithm for obstacle avoidance. For environmental representation, we apply feature point filtering in the scene to improve map accuracy. For more accurate pose estimation, we propose a stable feature-based localization algorithm that significantly reduces the Relative Pose Error (RPE). It mitigates the impact of noise-prone objects like leaves and grass and ensures reliable pose estimation in dynamic agricultural environments. Additionally, the PPM-Unet model is introduced for semantic segmentation, significantly outperforming baseline models such as Unet and MANet in segmenting low-lying obstacles. We also created the ASO3600 dataset to train the model. It contains data on small obstacles collected from diverse agricultural environments. We have validated our proposed navigation system in large-scale farms. The results show that our method achieves higher localization accuracy and better obstacle avoidance performance compared to existing methods. This work is available at https://github.com/YaoXingbo/StableAgriNav.