Season-robust localization for autonomous robot in orchard using two-stage LiDAR points segmentation with sensor fusion

PAN SIYU, Hu Yaohua, Ohya Akihisa, Ayanori Yorozu

A robust cross-season, multi-route localization framework is proposed for agricultural environments with significant seasonal variations. Based on intensity-calibrated maps constructed across four seasons, a two-stage point segmentation method is employed to extract geometrically salient features, which are then used as inputs for NDT-based scan matching. An extended Kalman filter (EKF) is further integrated to fuse IMU and odometry measurements, thereby improving localization stability. Experimental results obtained under different seasonal conditions and across three route types demonstrate that the proposed method achieves both high accuracy and real-time performance. Specifically, the method attains an absolute trajectory error (ATE) of <= 0.100 m, an absolute rotation error (ARE) of < 5 degrees, and average execution time of < 2.5 ms. The average ATE values across four seasons are 0.089 m, 0.092 m, and 0.091 m, while the corresponding average ARE values are 1.089 degrees, 1.132 degrees, and 1.218 degrees, and the average execution time cross four seasons are 1.755 ms, 1.134 ms, and 1.663 ms for structured, circular, and unstructured routes, respectively. Local evaluations during turning further demonstrate stable localization performance across both routes and seasons, with ATE ranging from 0.057 to 0.098 m and ARE ranging from 1.044 degrees to 3.709 degrees, indicating strong robustness under dynamic motion conditions. Compared with existing methods, the proposed framework significantly reduces localization errors and enhances robustness in seasonally varying agricultural environments.