Core Concepts
HyDRa introduces a novel camera-radar fusion architecture for robust depth prediction and state-of-the-art performance in 3D perception tasks.
Abstract
Abstract:
Low-cost vision-centric 3D perception systems have progressed, narrowing the gap with LiDAR-based methods.
HyDRa introduces a hybrid fusion approach combining camera and radar features for improved depth predictions.
Introduction:
Reliable 3D perception is crucial for autonomous vehicles to navigate dynamic environments.
Camera-LiDAR fusion is standard, but LiDAR's high cost drives interest in vision-centric systems.
Incorporating Radar Sensors:
Radar sensors offer resilience to adverse conditions and valuable velocity information.
Combining radar with cameras enhances 3D perception systems' potential.
HyDRa Architecture:
Modality-specific feature encoder processes multi-view images and radar point cloud data.
Height Association Transformer associates radar features with image features for robust depth predictions.
BEV Fusion module generates initial BEV representation by combining radar pillars channels with image features.
Experiments and Results:
HyDRa achieves state-of-the-art results in camera-radar fusion, surpassing previous methods in NDS and AMOTA scores.
Superior performance in occupancy prediction on the Occ3D benchmark highlights the effectiveness of dense representations.
Stats
HyDRaは64.2 NDSと58.4 AMOTAで公開されたnuScenesデータセットで新しいカメラレーダー融合モデルの最先端を達成しました。