面向自动驾驶的鸟瞰视图目标检测研究综述OA

Bird's-eye-view(BEV)representation,based on a unified spatial coordinate system,enables multi-sensor spatiotemporal alignment and feature fusion,demonstrating significant advantages in 3D object detection tasks for autonomous driving.Firstly,this paper elaborates on the benefits of BEV representation,and introduces commonly used autonomous driving datasets and evaluation metrics for 3D object detection.Differing from existing review lit-erature that focuses on input modality characteristics or modality feature fusion,this study systematically organ-izes current research works by using BEV representation density as the perspective,categorizing approaches into sparse representations,dense representations and hybrid sparse-dense representations.This study reveals the inher-ent trade-off dilemma between temporal and spatial overhead and representation completeness in existing methods.Finally,this study outlines future research trends toward low-cost and high-precision BEV object detection in autonomous driving,including but not limited to all-weather robust perception,camera-parameter-independent ad-aptive geometric modeling,and autonomous driving open-world object detection.These proposed research direc-tions aim to inspire BEV perception researchers and accelerate technological breakthroughs in autonomous driving systems.