一种新的基于单目视觉的广义障碍物检测方法

 障碍物检测是辅助驾驶、机器人导航等领域的核心问题之一.本文提出一种新的基于特征点道路面投影位移矢量的单目视觉广义障碍物检测方法.基于道路平面假设,利用特征点估计相机自运动参数,并利用此参数对相机的旋转运动进行补偿.利用逆透视投影变换,分别推导并证明了道路平面上的点和障碍物上的点的道路面投影位移矢量与相机位移矢量的关系.提出了一种区间统计方法,实现了相机位移矢量的鲁棒估计.最后,通过分析连续图像特征点的道路面投影位移矢量与相机位移矢量的关系,实现了广义障碍物检测.各种场景下的实验结果表明,该方法能够检测任意类型、形状的障碍物.与传统的运动补偿方法相比,具有更好的鲁棒性和准确性.

A New Method for Generalized Obstacle Detection Based on Monocular Vision

Obstacle detection is one of the key problems in driver assistance and robot navigation,etc.A monocular generalized obstacle detection method based on ground projection displacement vectors of feature points is proposed.The motion parameters of camera are estimated by feature points under the ground plane assumption and the rotation of camera is compensated by these parameters.The relationship between displacement vectors of camera and ground projection displacement vectors of both ground points and obstacle points are deduced by IPM (Inverse Perspective Mapping) respectively.An interval statistical method is proposed and the displacement vectors of camera are robustly estimated by this method.The generalized obstacle is detected by analyzing the relationship between ground projection displacement vectors of feature point pairs in image sequence and displacement vectors of camera.The experimental results under various scenes illustrate that the method can detect obstacle in arbitrary type and shape.Comparing with traditional obstacle detection method based on motion compensation,this method is more robust and accurate.