面向实景三维建模的倾斜影像像控点目标快速检索与高效刺点研究

A Research on Fast Retrieval and Efficient Marking Point of Photo-control point Object for 3D Modeling via Oblique Photogrammetry

本文针对无人机倾斜摄影测量处理过程中空三计算加密时需要通过人工进行像控点查找与刺点,从而导致工作效率低下的问题,研究像控点的形状特征和目标属性,采用YOLO算法进行目标检测与特征提取。针对YOLO算法只能处理1024×1024像素三通道图像的局限,本文将倾斜摄影测量影像进行像素分割后放入内存,使用416×416像素大小、正交方向步进值为208像素的滑动窗口截取小图,并采用循环调用目标检测框架的方法对小图进行预测,直到匹配到像控点目标为止,从而实现了像控点的快速检测和刺点目的。

Currently,the efficiency of manual image control point searching and marking is very low during aerial triangulation computingIn this paper,we focus on studying the shape features and target properties of image control points,and provide a new method to efficiently detect object and extracting feature in terms of the YOLOv3 deep learning algorithm.We propose to segment the oblique photogrammetric images into the memory,and use the sliding window with the pixel size of 416×416 as well as 208 pixel orthogonal direction step value to intercept the small imagebecause the image with 1024×1024 size and three channel cant be solved by YOLOv3.Finally,we design a program cycle call target detection framework to predict the small cropped image until it matches the image control point target,so as to realize the goal of fast detection and marking point.