Unsupervised learning for surveillance planning with team of aerial vehicles

In this paper, we extent an existing self-organizing map (SOM)-based approach for the Dubins traveling salesman problem (DTSP) to solve its multi-vehicle variant generalized for visiting target regions called k-DTSP with Neighborhoods (k-DTSPN). The Dubins TSP is a variant of the combinatorial TSP for curvature-constrained vehicles. The problem is to determine a cost efficient path to visit a given set of continuous regions while the path allows to satisfy kinematic constraints of non-holonomic vehicles. The k-DTSPN is a generalization to determine k such paths, one for each vehicle. Although the k-DTSPN has been addressed by evolutionary methods, the proposed approach is able to provide solutions very quickly in units of seconds on conventional computationally resources which makes the proposed SOM-based approach suitable for on-line planning. The studied problem is motivated by surveillance task in which it is required to quickly provide information about the given set of target locations. Therefore, real computational requirements are crucial properties of the desired k-DTSPN solver. The proposed method meets this requirement and feasibility of the found solutions are demonstrated not only in computer simulations but also with a practical deployment on real aerial vehicles.