Near-real-time gradually expanding 3D land surface reconstruction in disaster areas by sequential drone imagery
Near-real-time gradually expanding 3D land surface reconstruction in disaster areas by sequential drone imagery
复制标题
DOI：
10.1016/j.autcon.2021.104105
复制
发表时间：
2022-03-01
影响因子：
10.3
通讯作者：
Yamazaki, Fumio
中科院分区：
文献类型：
--
作者：
Cheng, Min-Lung;Matsuoka, Masashi;Yamazaki, Fumio
无人机进入灾区的能力已经被证明是强大而灵活的，可以为环境观测获取第一手光学图像数据。然而，这些影像数据通常经过后处理，三维（3D）产品主要用于精确的土地调查。后处理过程过于耗时，无法满足即时决策支持和救援响应要求。因此，本文拟开发一套系统的工作流程，利用无人机顺序获取的光学图像，实现灾区的实时三维重建。本研究提出了一种空间链接序列图像（SLSI）策略，用于图像定位和合适的立体对选择。此外，还建立了有效的外极立体对确定准则，使三维密集重建更加自动化和有效。在无人机捕捉新图像的同时，可以在计算机系统中逐步重建和扩展三维数字陆面。本文利用2016年日本熊本地震导致建筑物倒塌的图像数据集，模拟更有效的三维重建。尽管结果的准确性据报道接近一米，但在具有英特尔酷睿i5和16gb随机存取存储器（RAM）的iMac上执行所提出的方案时，每张图像的平均数据处理时间可以达到大约10秒的水平。因此，在灾难发生后不久支持紧急应用所需的效率和计算能力大大降低。
The ability of drones to access disaster areas has been proven powerful and flexible for acquiring first-hand optical imagery data for environmental observation. However, such imagery data usually undergo post processing, and the three-dimensional (3D) products are mainly for accurate land surveys. The postprocessing procedure is too time-consuming to meet instant decision support and rescue response requirements. Therefore, this paper intends to develop a systematic workflow that is able to achieve on-the-fly 3D reconstruction in disaster areas by optical imagery sequentially acquired by drones. This study proposes a strategy to spatially link sequential images (SLSI) for image localization and suitable stereopair selection. In addition, the criteria for valid epipolar stereoapair determination are developed to make the 3D dense reconstruction more automatic and effective. The 3D digital land surface can be gradually reconstructed and expanded in the computer system while the drone is capturing new images. This paper utilizes the imagery dataset of collapsed buildings induced by the 2016 Kumamoto earthquake in Japan to simulate the more effective 3D reconstruction. Although the accuracy of the consequence is reported to be closely one meter, the mean data processing time for every image can achieve the level by approximately ten seconds while performing the proposed scheme on an iMac with Intel Core i5 and 16 GB random access memory (RAM). As a result, the efficiency and computational power needed are significantly reduced to support emergency applications soon after a disaster occurs.