目的 目标检测是遥感智能解译中重要的研究方向之一,大多数目标检测算法难以实现密集排列的旋转目标的高精度检测。本文提出了一种基于关键点与引导向量预测的目标检测算法,实现高精度旋转目标检测的同时,还可对目标的朝向进行表征。方法 首先提出了一种新的旋转目标建模方式,将目标检测分解成中心点、头部顶点、引导向量以及目标宽度的参数回归以更贴合检测目标；其次设计旋转椭圆高斯核,能够更好的拟合遥感目标的形状,从而提升关键点的预测精度；最后通过预测中心点指向头部顶点的引导向量,完成同一个目标内中心点与头部顶点的匹配,从而生成一个精准的带方向的旋转矩形检测框。结果 在大长宽比舰船目标的HRSC数据集上的实验结果表明,相比于其它主流的目标检测算法,本文所提算法获得了更好的检测结果,平均精度分别达到了90.78%的平均精度(VOC 2007)和97.85%的平均精度(VOC 2012)。在小长宽比飞机目标的UCAS-AOD数据集上达到了98.81%的平均精度。结论 本文算法利用椭圆高斯核计算中心点与头部顶点,并设计引导向量对点匹配关系进行约束,实现了旋转目标的方向检测,实验结果表明了本文算法的可行性与有效性。

Objective Object detection is one of the important research directions in remote sensing intelligent interpretation,and most object detection algorithms are difficult to achieve high accuracy detection of densely arranged rotating objects. This paper proposes an object detection algorithm based on key points and guide vector prediction to achieve high accuracy detection of rotating objects while characterizing the orientation of the objects. Method Firstly, this method proposes a new way of directional object modeling, decomposing the object detection into center point, head vertex, guidance vector and object width parameter regression to better fit the detection object; secondly, designing a rotating elliptical Gaussian kernel, which can better fit the shape of the ship"s object, thus improving the prediction accuracy of key points; finally, by predicting the guidance vector from the center point to the head vertex, the matching of the center point and the head vertex in the same object is completed, thus generating an accurate rotating rectangular detection frame with direction. Result The experimental results on HRSC ship datasets show that compared with other mainstream object detection algorithms, the proposed algorithm achieves better detection results with an average accuracy of 90.78%(VOC 2007) and 97.85%(VOC 2012), respectively. The experimental results on UCAS-AOD flight datasets show that compared with other mainstream object detection algorithms, the proposed algorithm achieves better detection results with an average accuracy of 90.78%. Conclusion The algorithm in this paper proposed the elliptical Gaussian kernel to calculate the center point and head vertex, and designed a guide vector to constrain the point matching relationship, achieving direction detection of rotating targets. The experimental results show the feasibility and effectiveness of the algorithm in this paper.