Egocentric geolocation estimation in the Farm context via Monocular Vision and Variable Baseline

Fardad NasrollahzadehKhakiani, Mehran Mehrandezh, Kambiz Ghaemi Osgouie

Keywords: Agricultural robotics, Monocular Depth Estimation, MDE, Monocular visual odometry, ego-motion-geolocation

TL;DR: We present a monocular vision-based framework that overcomes scale ambiguity to achieve sub-meter geolocation accuracy for autonomous navigation in agricultural environments.

Abstract: Accurate geolocation is essential for autonomous navigation in agricultural environments, yet monocular depth estimation (MDE) suffers from scale ambiguity and drift. This paper introduces an egocentric geolocation framework that integrates monocular vision with a variable-baseline triangulation strategy. By anchoring triangulation to the first frame and incrementally expanding the baseline, the method enhances depth accuracy through increased parallax. ORB-SLAM2 is used for feature-based pose estimation, and triangulated scene coordinates are converted to global GPS positions via an ellipsoidal Haversine formulation. Tested on 1,857 monocular images from an agricultural dataset, the approach achieved sub-meter accuracy (MAE 0.3 m, RMSE 0.33 m), outperforming conventional fixed-baseline monocular odometry methods (>4 m error).

Submission Number: 64