创新背景

要辨别水中的黑斑是鲨鱼还是海草并不总是那么容易，在合理的条件下，无人机飞行员做出正确判断的几率通常只有60%。虽然这对公共安全有影响，但也可能导致不必要的海滩关闭和公众恐慌。

工程师们试图用人工智能(AI)提高这些“鲨鱼探测”无人机的准确性。尽管人工智能系统在实验室中显示出了巨大的前景，但在现实世界中很难正确使用，因此仍然是冲浪救生员无法触及的。重要的是，对这类软件的过度自信会带来严重的后果。

考虑到这些挑战，研究团队开始构建加强版鲨鱼探测器，并在现实条件下进行测试。通过使用大量的数据，为冲浪救生员创建了一个高度可靠的移动应用程序。

创新过程

自2016年以来，州政府一直在试验无人机作为鲨鱼发现工具，自2018年以来与新南威尔士州冲浪救生组织合作。训练有素的冲浪救生飞行员驾驶无人机在60米高的海洋上空飞行，在便携式屏幕上观看鲨鱼在水面下游动的形状的实时视频。

通过仔细分析良好条件下的视频片段来识别鲨鱼似乎很容易。但是，海水的清晰度、海面反光、动物深度、飞行员经验和疲劳都将实时探测的可靠性降低到预测的平均60%。当情况不明朗时，这种可靠性进一步下降。飞行员还需要准确地识别鲨鱼的种类，并分辨出危险和非危险动物的区别，比如经常被错误识别的鳐鱼。

人工智能驱动的计算机视觉一直被吹捧为一种理想的工具，可以在无人机播放的视频片段中虚拟“标记”鲨鱼和其他动物，并帮助识别海滩附近的物种是否值得关注。

此前人工智能鲨鱼探测系统的早期结果表明，这个问题已经得到解决，因为这些系统的探测精度超过90%。但要在新南威尔士州的海滩上扩大这些系统的规模，使其在现实世界中有所不同，一直具有挑战性。

研究人员的目标是通过一款新的鲨鱼探测手机应用程序来克服这些挑战。他们收集了大量无人机拍摄的视频，鲨鱼专家随后花了数周的时间检查视频，在数小时的视频中仔细追踪并标记鲨鱼和其他海洋动物。

使用这个新的数据集，研究人员训练了一个机器学习模型来识别十种海洋生物，包括不同种类的危险鲨鱼，如大白鲨和鲸鲨。然后他们将这个模型嵌入到一个新的移动应用程序中，它可以在无人机直播画面中突出鲨鱼，并预测物种。在现实条件下，新型探测器以一帧一帧的基础识别出了80%的危险鲨鱼。

创新关键点

研究人员收集了大量无人机拍摄的视频，鲨鱼专家随后花了数周的时间检查视频，在数小时的视频中仔细追踪和并记鲨鱼和其他海洋动物。

创新价值

人工智能可以在提高这些飞行效率、提高无人机监视的可靠性方面发挥关键作用，并可能最终导致全自动的鲨鱼发现操作和可信的自动警报。

新型应用程序不仅可以提高海滩的安全性，还可以帮助监测澳大利亚海岸线的健康。

Innovative use of AI to improve the accuracy of "shark detection" drones

The state government has been experimenting with drones as a shark spotting tool since 2016, in partnership with Surf Life Saving NSW since 2018. Trained surf lifesaving pilots fly drones 60 metres above the ocean, watching live video of shark shapes swimming below the surface on a portable screen.

It seems easy to identify sharks by carefully analyzing video footage of good conditions. However, the clarity of the water, surface reflection, animal depth, pilot experience and fatigue all reduced the reliability of real-time detection to an average of 60 percent of the prediction. This reliability deteriorates further when the situation is unclear. Pilots also need to accurately identify shark species and tell the difference between dangerous and non-dangerous animals, such as rays, which are often misidentified.

Ai-powered computer vision has been touted as an ideal tool to virtually "tag" sharks and other animals in video clips played by drones, and to help identify if species near beaches are of concern.

Earlier results from AI shark detection systems suggested that the problem had been solved, as these systems were more than 90 percent accurate. But scaling up these systems on NSW beaches to make a difference in the real world has been challenging.

Researchers aim to overcome these challenges with a new shark detection phone app. They collected plenty of drone footage, and shark experts then spent weeks examining the video, carefully tracking and tagging sharks and other Marine animals in hours of footage.

Using this new dataset, the researchers trained a machine learning model to identify ten species of Marine life, including different species of dangerous sharks such as great whites and whale sharks. They then embedded the model in a new mobile app that highlights sharks in live drone footage and predicts species. Under real-world conditions, the new detector identified 80 percent of the dangerous sharks on a frame-by-frame basis.