全卷积神经网络的车辆点云目标精细化检测

Vehicle detection from LiDAR point cloud using 3D fully convolutional network

针对无人驾驶技术高速发展中车辆目标的3D检测仍存在局限性的问题,该文提出了一种基于全卷积神经网络的车辆点云三维目标检测框架。进行了深度学习技术在二维图像的目标检测成熟应用的调查,将全卷积神经网络的目标检测扩展到三维点云数据。该算法在KITTI数据集上进行了测试,并与先前基于点云的车辆检测方法进行比较表明算法性能有着显著提高。研究结果可以应用于激光雷达点云实现车辆检测任务,从而可以较好地服务于自动驾驶。

With the rapid development of driverless technology,the detection of vehicle targets is becoming more and more important in the field of large-scale traffic surveying.In the context of the deep detection technology mature application in the target detection of two-dimensional images,we extend the target detection based on the full convolutional neural network to the 3 D point cloud data.This paper is based on the LiDAR point cloud to realize the vehicle detection task,which can better serve the automatic driving.The algorithm we proposed was tested on the KITTI dataset and compared with previous point cloud-based vehicle detection methods.The results show that our algorithm performance is significantly improved.