Robust BEV 3D Object Detection for Vehicles with Tire Blow-Out

Area of Science:

• Computer Vision
• Autonomous Systems
• Robotics

Background:

• Bird's-Eye View (BEV) methods are crucial for autonomous vehicle perception, offering advantages over LiDAR.
• Existing BEV methods fail during tire blow-outs due to reliance on accurate camera calibration, posing safety risks.

Purpose of the Study:

• To develop a robust BEV perception method for autonomous vehicles that can withstand tire blow-out scenarios.
• To address the limitations of current BEV methods when camera calibration is compromised by tire deflation.

Main Methods:

• Proposed a Geometry-Guided Auto-Resizable Kernel Transformer (GARKT) method specifically for tire blow-out situations.
• Developed a camera deviation model for tire blow-outs and utilized geometric priors with auto-resizable kernels.
• Encoded resizable perception areas and flattened them to generate BEV representations.

Main Results:

• GARKT achieved a nuScenes Detection Score (NDS) of 0.439 on a novel blow-out dataset.
• Maintained a robust NDS of 0.431 even with a completely flat tire, outperforming other transformer-based BEV methods.
• Demonstrated near real-time performance at approximately 20.5 frames per second on a single GPU.

Conclusions:

• The GARKT method significantly improves the robustness and safety of autonomous vehicle perception systems during tire blow-outs.
• GARKT offers a practical solution for maintaining reliable perception in challenging, real-world driving conditions.
• The proposed approach provides a promising direction for future autonomous driving safety systems.