The detection of fireballs streaks in astronomical imagery can be carried out by a variety of methods. The Desert Fireball Network uses a network of cameras to track and triangulate incoming fireballs to recover meteorites with orbits and to build a fireball orbital dataset. Fireball detection is done on-board camera, but due to the design constraints imposed by remote deployment, the cameras are limited in processing power and time. We describe the processing software used for fireball detection under these constrained circumstances. Two different approaches were compared: (1) A single-layer neural network with 10 hidden units that were trained using manually selected fireballs and (2) a more traditional computational approach based on cascading steps of increasing complexity, whereby computationally simple filters are used to discard uninteresting portions of the images, allowing for more computationally expensive analysis of the remainder. Both approaches allowed a full night’s worth of data (over a thousand 36-megapixel images) to be processed each day using a low-power single-board computer. We distinguish between large (likely meteorite-dropping) fireballs and smaller fainter ones (typical ‘shooting stars’). Traditional processing and neural network algorithms both performed well on large fireballs within an approximately 30 000-image dataset, with a true positive detection rate of 96% and 100%, respectively, but the neural network was significantly more successful at smaller fireballs, with rates of 67% and 82%, respectively. However, this improved success came at a cost of significantly more false positives for the neural network results, and additionally the neural network does not produce precise fireball coordinates within an image (as it classifies). Simple consideration of the network geometry indicates that overall detection rate for triangulated large fireballs is calculated to be better than 99.7% and 99.9%, by ensuring that there are multiple double-station opportunities to detect any one fireball. As such, both algorithms are considered sufficient for meteor-dropping fireball event detection, with some consideration of the acceptable number of false positives compared to sensitivity.

中文翻译：

---

沙漠火球网络数据处理管道对 CPU 处理要求最低的火球条纹检测

可以通过多种方法检测天文图像中的火球条纹。沙漠火球网络使用摄像机网络来跟踪和三角测量传入的火球，以恢复具有轨道的陨石并建立火球轨道数据集。火球检测是在机载摄像头上完成的，但由于远程部署所施加的设计限制，摄像头的处理能力和时间有限。我们描述了在这些受限情况下用于火球检测的处理软件。比较了两种不同的方法：(1) 具有 10 个隐藏单元的单层神经网络，使用手动选择的火球进行训练；(2) 一种更传统的基于复杂度增加的级联步骤的计算方法，从而使用计算简单的过滤器来丢弃图像中不感兴趣的部分，从而对剩余部分进行更昂贵的计算分析。这两种方法都允许每天使用低功耗单板计算机处理一整夜的数据（超过一千张 36 兆像素的图像）。我们区分大的（可能是陨石坠落的）火球和较小的较暗的火球（典型的“流星”）。传统的处理和神经网络算法在大约 30 000 张图像数据集中的大火球上都表现良好，真阳性检测率分别为 96% 和 100%，但神经网络在较小的火球上的成功率要高得多。分别为 67% 和 82%。然而，这种改进的成功是以神经网络结果的误报显着增加为代价的，此外，神经网络不会在图像内产生精确的火球坐标（因为它是分类的）。对网络几何的简单考虑表明，通过确保存在多个双站机会来检测任何一个火球，计算出三角大火球的总体检测率优于 99.7% 和 99.9%。因此，这两种算法都被认为足以检测流星坠落的火球事件，但与灵敏度相比，还需要考虑可接受的误报数量。对网络几何的简单考虑表明，通过确保存在多个双站机会来检测任何一个火球，计算出三角大火球的总体检测率优于 99.7% 和 99.9%。因此，这两种算法都被认为足以检测流星坠落的火球事件，但与灵敏度相比，还需要考虑可接受的误报数量。对网络几何的简单考虑表明，通过确保存在多个双站机会来检测任何一个火球，计算出三角大火球的总体检测率优于 99.7% 和 99.9%。因此，这两种算法都被认为足以检测流星坠落的火球事件，但与灵敏度相比，还需要考虑可接受的误报数量。