• Two-stage segmentation–based geometrically salient points combined with sensor fusion achieve accurate global translation and rotation localization across routes and seasons. • Local turning evaluations confirm stable localization during turning movements in multiple routes. • Compared with previous localization algorithms, the proposed method achieves higher computational efficiency (average execution time across four seasons < 2ms) and robustness under seasonal variations. 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°, 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°, 1.132°, and 1.218°, 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° to 3.709°, 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.
Pan et al. (Fri,) studied this question.