PRUNE: Selective LiDAR-Camera Projection via Multi-Stage Evidence Rejection for Semantic Fusion in Unstructured Forests

Abstract

Autonomous navigation in unstructured outdoor environments such as forests requires robust multimodal perception, where projecting Light Detection and Ranging (LiDAR) point clouds onto red green blue (RGB) camera images is a key step. Yet, geometrically valid projections can still be semantically unreliable because of horizon effects, depth discontinuities, and occlusions, causing incorrect image labels to be transferred to 3D points. We present PRUNE (Projection Rejection of Unreliable Neighborhood Evidence), a model-agnostic front end that selectively accepts, suppresses, or discards image evidence before semantic mapping. PRUNE validates temporal consistency, optionally compensates for ego-motion, and applies configurable gates for invalid-mask rejection, depth-edge rejection, occlusion consistency, and confidence thresholding. The method is training-free and requires no neural inference. We evaluate the three primary rejection gates on three calibrated forest sequences collected with an Ouster OS1-128 LiDAR and an Intel RealSense D435i RGB camera. Across all sequences, the full configuration retains 85.0 percent of projected points while rejecting 5.5 percent through the invalid-mask gate, 1.2 percent through the depthedge gate, and 9.2 percent through the occlusion gate. Pointlevel overlap analysis shows that the rejection mechanisms are largely complementary, with only 3.0 percent mean pairwise overlap. These results demonstrate that selective projection effectively identifies unreliable image evidence before it contaminates semantically persistent 3D maps.