Core Concepts
PreSight introduces a novel framework leveraging historical data to construct static priors, enhancing online perception in autonomous driving systems.
Abstract
1. Introduction:
Autonomous vehicles rely on perception systems for navigation and safety.
Humans use mental maps to supplement real-time perception in new environments.
2. Vision-Based Perception:
Dynamic object perception requires real-time observation.
Static environment perception can be pre-mapped and annotated.
3. PreSight Framework:
Utilizes historical traversal data to generate static priors without manual annotation.
Enhances online perception models with city-scale neural radiance fields (NeRF) priors.
4. Experimental Results:
Demonstrates significant improvements in HD-map construction and occupancy prediction tasks.
Shows compatibility with diverse online perception models on the nuScenes dataset.
5. Methodologies:
Building City-Scale NeRF involves partitioning cities into tiles and optimizing sub-fields for detailed coverage.
Prior Extraction uses a ray marching algorithm to identify occupied voxels and extract features for robust priors.
6. Comparison:
Outperforms other priors like historical point clouds and Neural Map Prior in HD map construction and occupancy prediction tasks.
7. Ablation Studies:
Distilled semantic features significantly improve performance in both HD map construction and occupancy prediction tasks.
Performance improves as more prior scenes are utilized, showcasing scalability of the framework.
Stats
Neural Radiance Fields (NeRF)を使用して過去のトラバーサルデータから静的先行メモリを構築します。
実験結果は、HDマップ構築と占有予測タスクで顕著な改善を示しています。