障碍物检测与车辆视觉定位一体化的方法研究

Research on the method of integrating obstacle detection and vehicle visual positioning

在自动驾驶的研究中,障碍物检测与车辆视觉定位一直是2个研究热点。针对在目前的自动驾驶解决方案中,障碍物检测与车辆视觉定位是被完全划分为2个模块单独完成的,部分计算得到的模块并未实现最佳的利用这一缺陷,提出了以立体匹配恢复出的视差图为节点,左右目影像间进行障碍物检测,前后帧之间以图像点匹配计算得到的位姿作为初值,辅助利用三角化方法恢复出点云,完成车辆的定位,构建了障碍物检测与车辆视觉定位的一体化系统,实现了对立体匹配结果的双重利用。在自动驾驶开源数据集上进行测试,障碍物检测具有较好的检测结果。里程计均方根误差为0.012766 m,全程轨迹误差平移百分比为0.4%,在精度上取得了较好的效果,且具有更强的鲁棒性。

In the research of autonomous driving,obstacle detection and vehicle visual localization have always been two research hotspots.In the current autonomous driving solution,obstacle detection and vehicle visual positioning are completely divided into two modules to complete separately,and some of the calculated modules are not optimally utilized.A stereo matching method is proposed.The recovered disparity map is a node.Obstacle detection is performed between the left and right eye images.The pose calculated by image point matching is used as the initial value between the front and rear frames.The triangulation method is used to restore the point cloud to complete the vehicle positioning.An integrated system for obstacle detection and vehicle visual positioning.The three-dimensional spatial position information recovered by stereo vision is combined with the image information to realize the dual utilization of the stereo matching results.Tested on the open source dataset for autonomous driving,obstacle detection has good detection results.The root mean square error of the odometer is 0.012766 m,and the translation percentage of the whole trajectory error is 0.4%,which has achieved good results in accuracy and has stronger robustness.