BEVHeight++: Toward Robust Visual Centric 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Current autonomous driving systems primarily rely on ego-vehicle sensors.
• Roadside cameras offer an alternative perception approach but face challenges with existing vision-centric methods.
• State-of-the-art methods struggle with roadside cameras due to depth recovery issues at increasing distances.

Purpose of the Study:

• To develop an effective perception method for autonomous driving using roadside cameras.
• To address the limitations of current vision-centric detection methods in roadside camera scenarios.
• To propose a novel approach, BEVHeight++, for improved 3D object detection and scene understanding.

Main Methods:

• Proposed BEVHeight++, a method that regresses height to the ground for a distance-agnostic formulation.
• Incorporated both height and depth encoding techniques for robust 2D to Bird's-Eye View (BEV) projection.
• Evaluated the method on popular 3D detection benchmarks for roadside cameras and ego-vehicle scenarios.

Main Results:

• BEVHeight++ significantly outperforms previous vision-centric methods on roadside camera benchmarks.
• In ego-vehicle scenarios, BEVHeight++ surpasses depth-only methods, showing notable improvements in NDS and mAP on nuScenes and nuScenes-C datasets.
• Consistent performance gains were observed across KITTI, KITTI-360, and Waymo datasets, including robustness to object-level distortion.

Conclusions:

• BEVHeight++ offers a simple yet effective solution for camera-only perception in autonomous driving.
• The proposed height regression method enhances accuracy and robustness in 3D object detection from roadside cameras.
• This approach extends perception capabilities beyond the ego-vehicle's visual range, paving the way for more comprehensive environmental awareness.