From Crop-Row Perception to Autonomous Row Following: A Review of Perception, Navigation, and Control for Agricultural Robots

Crop-row following is an important basis for inter-row operations such as sowing, weeding, spraying and harvesting. Its performance depends not only on the accuracy of crop-row perception but also on the coordination of navigation representation, localization and state estimation, path planning, headland turning, and obstacle handling. Accordingly, this review focuses on the research and analysis of autonomous tracking of agricultural robots. This review examines crop-row perception methods based on vision, LiDAR, and multimodal fusion. The applicability of semantic segmentation, object detection, geometric constraints, and structural priors is analyzed under complex lighting, weed interference, crop occlusion, and growth stage variation. The review further examines navigation representation methods, such as crop-row centerlines, keypoints, and drivable corridors, and analyzes state estimation problems in constrained environments such as coordinate transformation, multi-sensor fusion, SLAM, and GNSS. At the planning level, the focus is on discussing the collaborative relationship between row entry, reference path generation, headland detection and turning, obstacle perception, global coverage planning, and local replanning. This review compares the differences in accuracy, smoothness, real-time performance, kinematic constraints, and operational adaptability between different path tracking and motion control methods. This review provides a framework for understanding multimodal perception, task-constrained planning, and long-term autonomous operation in complex agricultural environments.