基于单目视觉的横穿障碍物检测

提出一种基于单目视觉的横穿障碍物检测方法.首先，基于道路平面假设，根据特征点的位置约束以及逆透视投影变换下的性质，提取地面特征点对.其次，采用迭代加权最小二乘法估计自车平移和旋转运动参数.然后，利用估计的运动参数对图像光流进行旋转补偿，并基于道路C 速度空间生成障碍物的候选标记点.最后，对候选标记点进行分组聚类和验证，确定横穿障碍物区域.不同交通场景下的实验结果表明，上述方法能够适用于各种自车运动，有效检测横穿障碍物.

Crossing Obstacle Detection Based on Monocular Vision

An algorithm for detecting obstacles crossing a vehicle's path was proposed by using a monocular camera. According to the flat road assumption, some pairs of feature points on the ground were extracted based on the feature points' position constraints and properties under IPM (inverse perspective mapping). Then, the ego-motion parameters were estimated from the feature point pairs using iteratively reweighted least squares algorithm. The estimated parameters were used to compensate the rotational motion in the optical flow field of the image, and the candidate points of the obstacle were marked in the road C-velocity space. Finally, an obstacle region was obtained by grouping and verifying these candidate points. The experimental results in various scenes illustrate that the method proposed is suitable for all kinds of ego-motion and can detect a crossing obstacle effectively.