Evaluation of Graph-Assisted RGB-D Processing and Vision–Attitude Fusion for Terrain-Slope Estimation in Hilly Farmland

Abstract

Reliable forward-looking slope estimation is important for agricultural machinery operating on hilly farmland. We evaluated a graph-assisted RGB-D processing pipeline comprising depth-edge enhancement, regional reliability filtering, point-cloud preprocessing, multi-stage RANSAC, and low-dimensional vision–attitude fusion. A controlled dataset contained 3520 cases covering 11 slopes, four terrain types, four depth-disturbance levels, and 20 repetitions. Under a shared success rule and ROI-geometric reference, the complete pipeline achieved 100% fit success, an MAE of 0.677 degrees, and an RMSE of 1.213 degrees; raw single-stage RANSAC achieved a lower MAE of 0.193 degrees. For 35 field records with reference slopes measured using a WT901C-232 sensor, visual estimation yielded an MAE of 2.459 degrees and an RMSE of 3.107 degrees. We also examined 739 deduplicated raw-aligned fusion records. Because their target labels were derived from LPMS-adjacent readings rather than an independent reference, we interpret those results only as fitting-consistency evidence. Conventional BP achieved an MAE of 0.288 degrees, compared with 0.291 degrees for self-attention BP, so the attention model showed no advantage. These results define the applicability limits and additional calibration requirements for field deployment.