基于地面先验的3D目标检测算法

Ground-aware 3D object detection algorithm

为了提高自动驾驶汽车环境感知的性能,增强单目相机对障碍物三维和边界信息的感知能力,提出了一种基于地面先验的3D目标检测算法。基于优化的中心网络(CenterNet)模型,以DLA(deep layer aggregation)为主干网络,增加目标3D边沿框中心点冗余信息预测。根据自动驾驶场景的地面先验信息,结合针孔相机模型,获取目标3D中心深度信息,以优化深度网络学习效果。使用KITTI 3D数据集评测算法性能,结果表明:在保证2D目标检测准确性的基础上,该算法运行帧率约20 fps,满足自动驾驶感知实时性要求;同时相比于CenterNet模型,在平均方位角得分(average orientation score)和鸟视图平均准确率(bird eye view AP)上分别有4.4和4.4%的性能提升。因而,该算法可以提高自动驾驶汽车对障碍物三维和边界信息的感知能力。

To improve the performance of environment perception of autonomous vehicles and enhance the ability of monocular cameras to perceive obstacle 3D and boundary information,this paper proposed a ground-aware 3D object detection algorithm.Based on an optimized CenterNet model,it used DLA as the backbone network to increase the redundant information prediction of 3D bounding box center.According to the ground-aware information of the autonomous driving scenario and combined with the pinhole camera model,it obtained the object 3D center depth information to optimize the deep network learning effect.This paper evaluated the algorithm performance using the KITTI 3D dataset.The results show that the algorithm runs at 20 fps on the basis of ensuring the accuracy of 2D object detection,which meets the real-time requirements of autonomous driving perception.At the same time,compared with the CenterNet model,there are performance improvements of 4.4 and 4.4%in the average orientation score and the bird eye view AP,respectively.Thus,the algorithm can improve the ability of autonomous vehicles to perceive obstacle 3D and boundary information.