Machine learning for drone detection from images: A review of techniques and challenges

Unmanned Aerial Vehicles (UAVs), popularly known as drones, have revolutionized many sectors. In precision agriculture, they are used to effectively sprinkle water, fertilizers, and pesticides. In cinematography, UAVs snap aerial images that were difficult or impossible to obtain in the past. However, just as they can be used for good, they also have the potential for malicious uses. For example, drug smugglers use drones to evade border surveillance and push their goods across countries. Additionally, in the hands of militants, drones can be used to launch ballistics on targets, leading to the loss of lives and properties. Thus, researchers have recently focused on designing automated tools to detect friendly from unfriendly drones. One effective tool for such is Machine Learning (ML). This paper reviews works that use ML to detect drones from images. The images include visible light, infrared, and thermal. After studying the papers, we present the taxonomy and trends in the field. In addition, we also provide open research issues: the development of lightweight models, the use of synthetic data, the adoption of auto-annotation models, and the employment of transformer-based models.