基于极曲线几何和支持邻域的鱼眼图像立体匹配

Stereo matching for Fish-eye images based on epipolar geometry and support neighborhood

提出了一种基于极曲线几何和变支持邻域的立体匹配算法来解决鱼眼立体视觉中图像变形导致的极曲线求取和匹配代价计算问题。首先,对鱼眼相机进行标定并获取相机的相关参数;针对鱼眼镜头的畸变问题,根据鱼眼镜头的优化投影模型推导出系统的极曲线方程,并利用得到的极线方程确定对应点的搜索范围。然后,根据同源像点在左右图像上的位置关系确定各中心像素点的支持邻域,并计算出不同视差条件下该支持邻域在另一幅图像上的对应支持邻域,利用获取的支持邻域计算出各点的匹配代价。最后,利用WTA(Winner Takes All)策略选取最佳匹配点得到最终的匹配结果。基于提出的极曲线和支持邻域对两组鱼眼图像进行了匹配实验并与传统方法进行了实验对比,结果表明:提出的方法的匹配准确度比传统方法分别提高了4.03%和4.64%。实验结果验证了极曲线的应用加快了匹配速度并减少了误匹配;支持邻域的使用使其对匹配代价计算的准确度优于传统方法。该算法满足了鱼眼图像立体匹配对信息获取速度、准确度和数量的要求。

A stereo matching algorithm based on epipolar geometry and support neighborhood was pro- posed to solve the problems of epipolar curve and matching cost calculation caused by image distortion in fisheye stereo system. Firstly, a fisheye camera was calibrated to obtain its relevant parameters. For the lens distortion of a fisheye lens, the epipolar curves of the system were derived according to the projection model of fisheye lens , and these epipolar curves were used to define searching scope when searching for corresponding points. Then, according to the position relationship between homol- ogous points on the left and right images, the support neighborhood of each pixel was determined, thecorresponding support neighborhood in other image was calculated in different parallaxes and the local matching cost was calculated based on support neighborhoods. Finally, matching results were ob- tained by using the Winner Takes All（WTA） strategy. The epipolar geometry and support neighbor- hood were used to perform a matching experiment and a comparison experiment for two groups of fish- eye images. Experimental results show that the matching accuracies of the two groups are increased by 4.03% and 4. 64% respectively as compared to traditional methods. It concludes that the epipolar curves speed up the matching speed and reduce the error; the accuracy rate of matching cost calcula- tion by proposed support neighborhood is also superior to that of the traditional matching for the fish- eye lens. This method meets the requirements of stereo matching of fisheye images for capturing speeds, accuracy and quantity.