激光雷达下车辆检测算法的研究

Research on Vehicle Detection Algorithms Based on LIDAR

在复杂的交通环境中,针对激光雷达应用于无人驾驶车辆的障碍检测的问题,提出了一种基于神经网络的前方车辆检测方法。首先,采用直通滤波算法对原始点云进行分割处理;其次,提出了一种端到端单阶段检测的深度神经网络,该网络利用空洞卷积对RetinaNet网络结构进行优化,增强该网络对车辆的准确性以及鲁棒性;最后,在KITTI数据集上进行了训练以及测试实验。结果显示,经过滤波处理后大幅度降低点云数量,更加精准的标记出检测范围。通过在测试KITTI数据集中测试不同检测算法处理结果对比得出,所提方法在精度提高的基础上,具有更快的检测速度,达到了预期效果,有较高的应用潜力。

In the complex traffic environment,aiming at the problem of the application of lidar to unmanned vehicle obstacle detection,a method of vehicle detection in front based on neural network was proposed.Firstly,the original point cloud is segmented by a straight-through filtering algorithm.Secondly,an end-to-end single-stage detection deep neural network is proposed.In this network,the structure of RetinaNet is optimized by using dilated convolution blocks to enhance the accuracy and robustness of the network against vehicles.Finally,training and testing experiments are carried out on the KITTI dataset.The results show that the number of point clouds is greatly reduced after filtering,and the detection range is marked more accurately.By comparing the processing results of different detection algorithms in the test KITTI data set,it is concluded that the proposed method has faster detection speed and higher application potential on the basis of improved accuracy.