BETAV: A Unified BEV-Transformer and Bézier Optimization Framework for Jointly Optimized End-to-End Autonomous...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous driving systems require advanced perception and planning for complex environments.
• Existing systems often struggle with accurate 3D perception and smooth trajectory generation.
• Challenges include integrating multi-view data and ensuring kinematic feasibility.

Purpose of the Study:

• To introduce BETAV, a novel framework for end-to-end autonomous driving.
• To address limitations in 3D perception accuracy and trajectory smoothness.
• To enhance scene understanding and motion planning for autonomous vehicles.

Main Methods:

• Utilizing Vision Transformers (ViTs) for multi-view camera data encoding into Bird's Eye View (BEV) representation.
• Employing a Bézier curve-based planning decoder for compact, continuous, and parameterized trajectory generation.
• Incorporating constraints for vehicle kinematics, obstacle avoidance, and directional alignment.

Main Results:

• Achieved state-of-the-art performance on the Nuscences benchmark datasets.
• Demonstrated superior robustness and generalization in diverse driving scenarios (Bench2Drive).
• Improved 3D perception accuracy and trajectory smoothness compared to existing methods.

Conclusions:

• The BETAV framework offers a unified approach to enhance autonomous driving perception and planning.
• Bézier-optimized planning ensures smoother, safer, and computationally efficient trajectories.
• The proposed method shows significant potential for real-world autonomous navigation.