TriQuery-BEV: Enhancing 3D Perception for Autonomous Driving with Temporal Query Filtering and Uncertainty-Aware...

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Existing bird's-eye-view (BEV) perception methods struggle in dynamic traffic scenarios due to depth-noise amplification, representation discontinuity from occlusion, and temporal drift.
• These limitations hinder the reliability of autonomous driving systems in complex, real-world conditions.

Purpose of the Study:

• To propose TriQuery-BEV, a novel modular framework designed to enhance BEV perception models.
• To specifically address geometric ambiguity, occlusion robustness, and temporal consistency in BEV query modeling for improved performance.

Main Methods:

• Introduced Query Mask (QM) for structured regularization within the BEV query space.
• Developed depth-modulated hybrid positional encoding (DM-HPE) for geometry-aware positional representations.
• Implemented a Temporal Query Filter (TQF) for uncertainty-aware temporal fusion of features.

Main Results:

• TriQuery-BEV demonstrated consistent improvements over the baseline BEVFormer on the nuScenes benchmark across various model scales.
• Achieved significant gains in nuScenes detection score (NDS) and mean average precision (mAP), with improvements up to 6.5%.
• Reduced key error metrics, including mean translation error (mATE), mean orientation error (mAOE), and mean velocity error (mAVE), by up to 15.0%.

Conclusions:

• The proposed TriQuery-BEV framework effectively enhances BEV perception capabilities in dynamic traffic scenes.
• The integrated components (QM, DM-HPE, TQF) contribute to improved robustness, geometric accuracy, and temporal consistency.
• TriQuery-BEV offers a computationally tractable solution for advancing the performance of autonomous driving perception systems.