Twenty-First Century Scarecrow

One of the unpleasant facts of city life is that pigeons make their homes wherever rooftops, ledges, or overhangs are found. These pesky creatures are more than just an annoyance to the people being hounded for scraps while eating outdoors — they also carry diseases, and their highly corrosive droppings cause irreparable damage to r​​ooftops, walkways, walls, vehicles, and just about everything else. This damage is estimated to be responsible for over one billion dollars in economic losses each year in the United States alone.

Traditional approaches to keep pigeons away rely on physical measures, such as plastic or metal spikes that are attached to horizontal surfaces. Of course there are huge numbers of such surfaces in a city, so it is not practical to deploy these measures everywhere that they are needed. A team at EPFL in Switzerland has recently designed a new method to keep pigeons at bay. Their approach involves the use of an autonomous drone that is capable of locating flocks of the birds, then flying to their location to scare them away. By using this tactic, it is possible for a single device to keep pigeons away from a relatively large area.

To keep costs low, a single, monocular pan-tilt-zoom camera was installed on a rooftop. The images captured by the camera are fed into a Faster region-based convolutional neural network (R-CNN). This network is a two-stage object detector, in which the first stage identifies regions of the image where the objects are located, after which the second stage draws bounding boxes around objects in these regions. The R-CNN uses Inception-ResNet v2 as a backbone, and was pretrained on the ImageNet dataset. The model was then modified such that it can recognize two classes (pigeon, other) and was fine-tuned by training it on data of pigeons in an urban environment.

Having identified pigeons in an image, the next step was to translate that location into GPS coordinates, such that the birds can be localized in the real world. Having chosen a monocular camera made this step more difficult, as there are no great solutions for recovering depth information from such a setup. The researchers took a bit of a shortcut, and assumed that all pigeons are of a fixed height. By using that information, in conjunction with the bounding box height as a percentage of the image height, a simple equation can determine the distance between the camera and the pigeon.

An off-the-shelf Parrot Anafi drone was then programmed to autonomously fly to the calculated GPS coordinates. In a series of trials, it was found that larger groups of pigeons were scared away simply by the takeoff of the drone (which was 40-60 meters away), whereas lone pigeons let the drone come within a few meters before they flew off. The team also found that the amount of time the drone remained in the area was important. Some pigeons would attempt to return to their previous spot almost immediately, but if the drone remained in the area for a time, it would keep them away. The researchers are planning to collaborate with zoologists as they continue to develop the device so that they can understand pigeon behavior more fully, and adapt their algorithms accordingly for maximal performance.