基于多模态鸟瞰视角的三维目标与三维车道线检测模型  

3D Object and 3D Lane Detection Model based on Multimodal Bird's-eye-view

作　　者：陈大宇 徐高伟 CHEN Dayu;XU Gaowei

机构地区：[1]智马达(上海)机器人科技有限公司,上海200232

出　　处：《上海汽车》2024年第9期29-36,共8页Shanghai Auto

摘　　要：自动驾驶车辆对动态物体的感知以及对静态车道线的检测往往需要满足高准确率和低延时的特征。针对不同的传感器特性(如激光雷达、摄像头),文章提出一种高效的多模态鸟瞰视角下(BEV,Bird'seye-view)的三维目标与三维车道线检测模型,称之为BOLD。该模型结合激光雷达定位准确度高以及图像语义信息丰富等互补性的优点,基于Lift-splat-shoot相机视椎模型将图像的二维特征转化至BEV空间,紧接着针对不同的任务共享骨干网进行特征提取,进一步降低模型成本,实现端到端的模型训练和低延时模型推理。Autonomous vehicles'perception of dynamic objects and static lane lines often re⁃quire high accuracy and low latency.According to different sensor characteristics(lidar,camera),an efficient multi-modal Bird's-eye-view(BEV)3D Object and 3D Lane Detection model(BOLD)is proposed.This model,which combines the complementary advantages of lidar's high positioning accu⁃racy and rich image semantic information,converts the 2D features of the image into the BEV space based on Lift-splat-shoot camera frustum model.The shared backbone network performs feature extrac⁃tion to further reduce model overhead and achieve end-to-end model training and low-latency model inference.

关 键 词：自动驾驶 环境感知 多模态 目标检测 车道线检测 

分 类 号：U463.6[机械工程—车辆工程]