VDVI与AKAZE结合的无人机森林影像匹配方法

Forest image matching method of unmanned aerial vehicle based on the combination of VDVI and AKAZE

【目的】无人机森林航摄影像的主体内容为颜色较单一的林冠,其纹理重复率高,而且存在非刚体性变换等导致影像的匹配难度高于普通测绘无人机航摄影像。如果直接采用普通特征算法进行影像匹配时,经常出现提取的特征点数量少、误匹配率高、特征点分布不均匀等问题。本研究探索了一种可见光植被指数(VDVI)与AKAZE结合的无人机森林影像匹配方法,以提高森林影像匹配的准确率。【方法】在森林航摄影像特征提取方面表现较优的AKAZE基础上,借助可见光植被指数(VDVI),在特征检测时选择了能够有效平滑区域内部同时保留边界信息的g3函数。特征检测时利用VDVI值进行检测,特征描述时采用原始影像信息与VDVI值相互校正的方法进行描述。【结果】在同一台I7-6700HQ处理器、8G内存的ALIENWARE工作站,建立Python 3.7和OpenCV 3.4环境,对内蒙古大兴安岭伊尔施林场的白桦和兴安落叶松林地无人机森林航摄影像进行了SIFT、SURF、普通AKAZE、颜色不变量AKAZE、可见光植被指数(VDVI)AKAZE等5种方法的对比分析。结果表明,VDVI与AKAZE结合的无人机森林影像匹配方法在无人机森林航摄影像上的平均匹配准确率为74.59%,比普通AKAZE高出了9.48个百分点;比颜色不变量AKAZE高出了5.26个百分点;其匹配速度在普通AKAZE和颜色不变量AKAZE之间。【结论】森林航摄影像主体内容的独特性,决定了采用普通测绘摄影测量和计算机视觉的特征算子进行影像匹配的效果不好。在AKAZE特征各向异性的非线性滤波的基础上选择合理的传导函数,配合可见光植被指数可有效降低阴影、反光、非林冠地物等的影响,进而改善了森林影像匹配的准确率,对研究和学习森林航摄影像处理具有重要意义。

【Objective】The forest aerial photography of an unmanned aerial vehicle(UAV)focus on crown canopy with relatively single color,resulting in the high repetition rate of texture.Moreover,due to the non-rigid transformation,it is more difficult to match images compared to aerial photography of UAV during ordinary mapping.Image matching by using ordinary feature algorithm directly often leads to some problems,such as a small quantity of extracted feature points,high mismatching rate,uneven distribution of feature points.In this study,a forest image matching method of UAV based on the combination of VDVI and AKAZE was explored to increase the accuracy rate of forest image matching.【Method】Based on AKAZE which performs well in feature extraction from forest aerial photography,the g3 function that can effectively smoothing inside of a region and maintain the boundary information during feature detection was chosen.Features were tested by VDVI value and were described by mutual calibration between the original image information and VDVI value.【Result】Python 3.7 and OpenCV 3.4 environments were established in the ALIENWARE stations with the same I7-6700HQ processor and 8G memory.Performances of SIFT,SURF,ordinary AKAZE,color invariant AKAZE and VDVI AKAZE in forest aerial photography of UAV of white birch and the Xing’an larch forest land in The Great Khingan Norbert Irsch forest land of Inner Mongolia were compared.According to results,the mean matching accuracy of the proposed method on forest aerial photography of UAV reaches 74.59%,which is 9.48%higher than that of ordinary AKAZE and 5.26%higher than that of color invariant AKAZE.Moreover,its matching speed is between those of ordinary AKAZE and color invariant AKAZE.【Conclusions】Due to the uniqueness of subjects of forest aerial photography,the image matching effect is unsatisfying when ordinary photogrammetry and feature operators of computer vision are used.Reasonable conduction function was chosen based on the nonlinear filtering of anisotropism of AKAZE features.The effects of shadow,reflected light and non-canopy ground features can be decreased effectively by combining with VDVI,thus improving the forest image matching accuracy.Research conclusions have important significance to study and learn forest aerial image processing.