Aerial Object Detection for Water-Based Search & Rescue

Responding to a water rescue situation is challenging. First responders need access to data as quickly as possible to increase the likelihood of a successful rescue. Using aerial imagery systems is especially useful in a search and rescue scenario because it provides a higher dimensional view of the search environment. Unmanned aerial vehicles can be easily used to acquire aerial image data. During water-based search and rescue scenarios, first responders sometimes deploy an inflatable marker called a rescue danbuoy. The danbuoy is fitted with a small conical sack known as a drogue, this ensures that the marker is not blown off course by the wind and instead follows the flow of the body of water. Tracking the danbuoy as it moves is of utmost importance in a water rescue. We present a new data-set “VisBuoy” with imagery containing instances of danbuoy markers and boats in real-world water-based settings. We also show how using various deep learning-based computer vision techniques, we can autonomously detect danbuoy instances in aerial imagery. We compare the performance of four state-of-the-art object detectors Faster RCNN Retinanet, Efficientdet and YOLOv5 on the “VisBuoy” data-set, to find the best detector for this task. We then propose a best model with a precision score of 74% which can be used in search and rescue operations to detect inflatable danbuoy markers in water-based settings.