基于全卷积神经网络和点云的道路检测算法

Road Detection Algorithm Based on Fully Convolution Network and Point Cloud

传统利用视觉传感器进行道路检测的算法受到环境光影响明显,难以在实际驾驶环境进行应用.本研究基于全卷积神经网络和激光雷达点云提出了一种道路检测算法,通过将点云进行投影,栅格化计算统计学特征,将无序的点云转化为全卷积神经网络能够训练的数据,进而对车辆前方的道路进行检测.在KITTI数据集上的训练和测试表明,该算法在识别精度、计算耗时、感知范围和稳定性上能够满足对于车辆前方道路的检测要求,具有鲁棒性好、全天应用的特点.

The traditional road detection algorithm based on visual sensors is obviously affected by ambient light and is difficult to be applied in actual driving environment.In this study,a road detection algorithm is proposed based on the fully convolutional neural network and LiDAR point cloud.By projecting the point cloud and rasterizing the statistical characteristics,the disordered point cloud is transformed into the data that the fully convolutional neural network can train,so as to detect the road in front of the vehicle.Training and testing on KITTI data sets show that the algorithm can meet the requirements of road detection in front of vehicles in recognition accuracy,calculation time,perception range and stability,and has the characteristics of good robustness and all-day application.