Hierarchical Taxonomic Abstraction for the Safe Handling of Novel Objects in Autonomous Driving Perception

Engineering notes

Key topics: autonomous driving, perception. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Object detectors for autonomous driving classify against a flat list of categories. When a scene contains an object that does not fit any of those categories (an overloaded rural truck, a horse trailer, a horse-drawn carriage), a flat classifier must either force a confident but wrong label or drop the object entirely. Both outcomes are unsafe. Building on a semantic-model framework for constraining pattern recognition, we replace the flat list with a hierarchical taxonomy and add a runtime abstraction mechanism: detections descend the taxonomy only as far as the visual evidence justifies and otherwise fall back to a coarser, still safety-relevant category. A per-branch safety floor bounds this fallback so the system never collapses into a useless generic bucket; below the floor it emits an explicit, localized unknown obstacle. Using pretrained YOLO boxes and CLIP zero-shot leaf scoring (no training), we evaluate a fair head-to-head against a flat baseline on the SegmentMeIfYouCan Road Anomaly set. On novel objects the hierarchy handles 100% safely (abstracted or flagged) versus 0% for the flat baseline, while on known objects it remains at least as specific (77% vs. 47% useful) and never makes a confident wrong claim.

Links and sources

Comments