Vision-Based Heading Correction for GPS-Resilient Autonomous Orchard Navigation

Precise localization is critical for enabling autonomous operations in precision agriculture, especially in structured environments such as orchards, where GNSS-based systems often fail due to canopy occlusion and signal degradation. While well-known sensor fusion techniques using IMU measurements improve performance, accumulated drift degrades fusion quality. This work proposes a novel vision-based measurement paradigm that provides local observations on the heading of a UGV purely from monocular RGB images, enabling autonomous navigation between crop rows in GNSS-degraded environments. Relying on a YOLOv11-seg semantic segmentation network, the system identifies the drivable area between tree rows and provides an image-based measurement of the lane center. This quantity is mapped, via a pinhole camera model, to a local orientation measurement that is explicitly modeled as a function of bias and noise. An Extended Kalman Filter then fuses this orientation observation with noisy global pose information and control inputs to produce a drift-compensated heading estimate, which is fed to a PID controller for closed-loop navigation. Simulation studies show that the system effectively compensates for both Gaussian noise and bias in the orientation measurements. Particular emphasis is placed on drift mitigation, which experimental results identify as a major challenge in real-world operation.