The surveillance of large areas to ensure local security requires remote sensors with high temporal and spatial resolution. Captive balloons with infrared and visible sensors, like ALTAVE captive balloon system, can perform a long-term day–night surveillance and provide security of large areas by monitoring people and vehicles, but it is an exhaustive task for a human. In order to provide a more efficient and less arduous monitoring, a deep learning model was trained to detect people and vehicles in images from captive balloons infrared and visible sensors. Two databases containing about 700 images each, one for each sensor, were manually built. Two networks were fine-tuned from a pretrained faster region-based convolution neural network (R-CNN). The network reached accuracies of 87.1% for the infrared network and 86.1% for the visible one. Both networks were able to satisfactorily detect multiple objects in an image with a variety of angles, positions, types (for vehicles), scales, and even with some noise and overlap. Thus a faster R-CNN pretrained only in common RGB (red, green, and blue) images can be fine-tuned to work satisfactorily on visible remote sensing (RS) images and even on the infrared RS images.

中文翻译：

---

使用深度学习的俘获气球图像对象检测系统

为了确保本地安全而进行的大范围监视需要具有高时空分辨率的远程传感器。具有红外和可见传感器的俘虏气球，例如ALTAVE俘虏气球系统，可以执行长期的昼夜监视，并通过监视人员和车辆来提供大范围的安全性，但这对人类来说是一项详尽的工作。为了提供更高效，更轻松的监控，训练了深度学习模型，以检测来自俘获气球红外和可见传感器的图像中的人和车辆。手动建立了两个数据库，每个数据库包含约700张图像，每个传感器一个。从预训练的基于区域的快速卷积神经网络（R-CNN）对两个网络进行了微调。该网络的红外网络的准确度达到87.1％，可见网络的准确度达到86.1％。这两个网络都能够令人满意地检测到图像中具有多个角度，位置，类型（用于车辆），比例尺甚至带有一些噪声和重叠的多个对象。因此，可以对仅在普通RGB（红色，绿色和蓝色）图像中预训练的更快的R-CNN进行微调，使其在可见遥感（RS）图像甚至红外RS图像上都能令人满意地工作。