Base Feature Point Recognition of Ratooning Rice Based on Improved YOLOv8s and Binocular Stereo Vision

Machine vision is key to agricultural machinery perception and decision-making. Accurate identification and 3D positioning of dormant buds of ratooning rice are critical for precise stubble retention and yield enhancement. Current solutions suffer from high small-target missed detection, inadequate stubble accuracy (monocular lacks 3D coordinates), and poor field robustness. This paper designs a system integrating improved YOLOv8s and an optimized binocular module. With anchor frame optimization and attention module enhancement, small-target mAP@0.5 reaches 91.7%. Optimized SGBM with IMU compensation achieves ±2.1 cm 3D positioning accuracy; field data enhancement boosts robustness by 35%. At 29.1 FPS (meeting real-time needs), it outputs 3D coordinates and recommended stubble height, breaking perception bottlenecks to support intelligent agricultural machinery decisions and ensure yield.