A Season‐Robust Long‐Term Localization Method Using Trunk Semantic Features in Dynamic Orchard Environments

ABSTRACT Localization is pivotal for enabling mobile robots to perform precision agriculture in orchards. However, dynamic factors such as seasonal variations present significant challenges for achieving long‐term localization. In dynamic orchard environments, tree trunks remain static and unaffected by seasonal changes, making them ideal landmarks for localization in dynamic scenarios. Therefore, we propose a novel season‐robust long‐term localization method using trunk semantic features. Firstly, we introduce a semantic extraction module that integrates camera and LiDAR data to capture tree trunk semantic features within a simultaneous localization and mapping (SLAM) framework, thereby generating lightweight static maps that remain invariant to seasonal fluctuations. Upon re‐entering the orchard environment, the robot utilizes a particle filtering framework to fuse LiDAR and GNSS data for long‐term localization based on the lightweight map. This process involves performing Normal Distribution Transform (NDT) registration on the original point cloud to predict particle states, comparing the trunk point cloud with the lightweight map to compute particle weights from LiDAR measurements, and optionally integrating GNSS measurements to enhance localization accuracy. Finally, we conducted extensive field data collection over a prolonged period in a real orchard environment, encompassing experimental data from both winter and summer. The performance of our proposed localization method, along with its capability to mitigate the effects of seasonal variations, was validated through localization experiments conducted across different seasons. Quantitative results indicate that the average error of the fused localization is 0.16 m, and the root mean square error (RMSE) is 0.18 m.