Neural Network Based Algorithm for Multi-UAV Coverage Path Planning

Unmanned Aerial Vehicles, better known as drones, have become an eye in the sky to aid men with a perspective from above. They provide real-time, high-resolution imagery at low cost. Originally thought for military applications, UAVs have found their way into mainstream usage thanks to the enhanced levels of safety and efficiency they bring. Robotic UAVs operate without an onboard pilot; the trend is of continuous evolution with an ever increasing level of automation. On the other side, Artificial Neural Networks lead the cutting-edge machine learning techniques, whose purpose is to render machines more and more intelligent by means of bio-inspired models. This work combines both the potential of artificial intelligence with drone-based surveillance capabilities - a fleet of AI-driven UAVs executes the Coverage Path Planning of a complex-shaped urban areas. The outcome is a strategic selection and planning of the trajectories over a map - while accounting for congestion, collisions, and image overlapping issues. The decision-making process is delivered by a balanced “explicit vs implicit” programming. The algorithm relies on a mixed-use of decentralized Artificial Neural Networks which confers elementary cognitive skills to each UAV, and a modified version of the famous A* pathfinder. Moreover, the training session of the Neural Network completely bypasses common drawbacks such as the need of large labeled databases or high computational resources. Particular attention is given to the scenario developed in the occasion of the Leonardo’s Drone Contest, proposed by the homonymous company. Further case studies focus on real urban areas, for which the grid resolution of the traditional Coverage Path Planning approaches can’t model the problem with sufficient accuracy.