Unsupervised learning for surveillance planning with team of aerial vehicles

Unsupervised learning‐based flexible framework for surveillance planning with aerial vehicles

“…The third class of approaches is evolutionary methods that can provide high‐quality solutions but are usually computationally very demanding. Finally, the fourth class is the recently proposed unsupervised learning which combines a solution of the sequencing part of the problem with the online sampling of the suitable heading values (Faigl & Váňa, ) and for the DTSPN also the waypoint locations (Faigl & Váňa, ). Selected approaches of the particular classes are briefly described in the rest of this section to support our selection of the considered methods in our effort towards a suitable solution for a practical deployment motivated by MBZIRC 2017.…”

“…The SOM‐based algorithm (Faigl & Váňa, ) has been significantly improved in (Faigl & Váňa, ), where the reported required computational time for scenarios (motivated by MBZIRC 2017) with 22 targets is found in less than 600 ms, while the solutions are better than those provided by the memetic algorithm (Zhang et al, ) with the computational time restricted to 10 s. Moreover, the SOM‐based approach has been generalized for the DTSPN, where the particular waypoint locations are determined during the winner selection together with the expected heading at the waypoint. In addition, the ‐DTSPN is addressed by creating an individual neural network for each vehicle, and during the winner selection, neurons from the network which represents a shorter tour are preferred to address the minmax variant of the ‐TSP (Somhom, Modares, & Enkawa, ).…”

“…Regarding approaches for the ‐DTSPN, they are similar to the ‐TSP in many ways (Bektas, ; Oberlin et al, ), especially the transformation/sampling‐based solvers, but only a few approaches directly address the challenges of the minmax variant of the ‐DTSPN. One of them is the memetic algorithm (Zhang et al, ), which has been compared with the additional direct approach based on SOM in Faigl & Váňa (). Another evolutionary based approach to the minmax variant of the ‐DTSPN has been proposed in Macharet et al (), but the authors do not report on the computational requirements, which also hold for the improved version presented in Macharet, Monteiro, Mateus, & Campos ().…”

“…Regarding the particular approaches to the ‐DTSP(N), we consider a purely combinatorial optimization approaches already proposed in the literature to address the ‐DTSPN and minmax variant of the ‐TSP. We also consider our previous effort towards surveillance planning with UAVs based on unsupervised learning of the self‐organizing map (SOM) first deployed in a solution of the DTSP in (Faigl & Váňa, ) and later generalized for the ‐DTSPN in (Faigl & Váňa, ). Following the sampling‐based approaches of the continuous optimization problem of Dubins planning (Oberlin et al, ; Obermeyer, Oberlin & Darbha, ), the variable neighborhood search (VNS) metaheuristic (Soylu, ) is also considered for a direct solution of the minmax ‐DTSP and its generalization to the ‐DTSPN.…”

“…Even though the presented work is built on the previous approaches published in the literature, that is, the VNS for the ‐TSP (Soylu, ) and SOM‐based unsupervised learning for the ‐DTSPN (Faigl & Váňa, ), they have been further developed to address the ‐DTSPN by the VNS‐based approach and the SOM‐based approach has been generalized to 3D surveillance planning. Therefore, we consider the main contributions of the paper with respect to the existing approaches as follows: Deployment of the VNS‐based ‐TSP solver in the ‐DTSPN. Fast and efficient initialization for the VNS‐based optimization in the ‐DTSPN. Comprehensive evaluation of the proposed VNS‐based solver and the existing memetic and SOM‐based approaches in the mbzirc22 scenarios of the ‐DTSPN with varying number of vehicles. Experimental verification of the found trajectories using real UAVs utilized in MBZIRC 2017. Generalization of the SOM‐based solver to 3D surveillance planning. Verification of the feasibility of the found 3D trajectories using real aerial vehicles. Since the developed unsupervised learning‐based solver allows a straightforward extension from the Dubins vehicle model to a Bézier curve or any similar model (e.g., Dubins‐Helix model; Wang, Wang, Tan, Zhou, & Wei, ), while the main principles are the same, we consider the proposed planner as a suitable flexible framework for surveillance planning with aerial vehicles. Unsupervised learning framework for surveillance planning addressing the ‐DTSPN but also the multivehicle planning problem where it is requested to quickly find surveillance trajectories considering the maximal vehicle velocities and acceleration limits that better fit the real motion capabilities of multirotor UAVs than Dubins vehicle describing curvature‐constrained trajectories suitable for fixed‐wing vehicles.…”