Tree Stem Perimeter Estimation for Forestry Robots via Residual Learning and LiDAR–RGB Fusion

Abstract

Accurate estimation of tree stem perimeter is essential for robotic forest mapping and inventory applications. While light detection and ranging (LiDAR)–camera fusion enables automated stem detection and geometric reconstruction, perimeter estimates derived from partial LiDAR observations exhibit systematic bias due to occlusions, limited angular coverage, and violations of ideal cylindrical assumptions. These effects introduce systematic errors that cannot be fully resolved by geometric reconstruction or temporal smoothing alone. This paper proposes a residual learning framework that enhances geometric perimeter estimation by learning a data-driven correction term while preserving the interpretability of the underlying model. A mobile robot equipped with a three-dimensional LiDAR and an RGB camera collects synchronized data in forest environments. Tree stems are detected in RGB images and associated with LiDAR points through calibrated projection. A geometric baseline perimeter is computed by fitting a circular model to a diameter-at-breast-height cross-section extracted from incomplete LiDAR observations. A shallow multilayer perceptron then predicts the residual between the geometric estimate and manually measured ground-truth perimeter using observation-derived features. Filtering improves stability but does not remove systematic bias, whereas residual learning achieves consistent bias correction. Experimental results demonstrate reductions exceeding 40% in both mean absolute and root mean squared errors, together with a substantial improvement in the coefficient of determination from 0.48 to 0.86. Error distributions become more centered and consistent across varying sensing distances and stem sizes, confirming robust generalization to unseen trees. The proposed method operates as a lightweight post-processing module, making it suitable for real-time deployment on mobile forestry robots.