基于自监督学习的越野道路检测算法

摘要

An algorithm focused on detecting off-road is presented.Drivable area is detected by lidar as sample of the current road.Navigation and position system is applied to track the nearby drivable area to locate it in the history images.After that the pixels in these areas are abstracted to train road models in different distances,which are subsequently used to evaluate the likelihood of different areas in the current image.Finally,a RANSAC algorithm is applied to calculate the optimal curve of road center.The result of experiments indicates that this algorithm is robust because it can adapt to changing road conditions.It takes about 97 ms to process one image on average after hardware accelerating.%提出了一种针对越野道路的检测算法：利用激光雷达检测出近处的可通行区域，作为当前道路的样本；利用导航定位系统追踪近处可通行区域在历史图像中的对应区域，提取这些区域内的像素值，训练不同距离下的道路模型，用其评估当前图像中各个区域的道路似然度；最后采用 RANSAC 算法计算最优的道路中心曲线。试验结果表明：该算法能根据道路变化调整道路模型，鲁棒性强，经过硬件加速后平均每帧图像的处理时间约为97 ms。