UAV path planning based on event density detection

Transcription

1 UAV path planning based on event density detection Roberto Henriques, Fernando Bação and Victor Lobo,2 Instituto Superior de Estatística e Gestão de Informação, ISEGIUNL 2 Portuguese Naval Academy {roberto, bacao, vlobo}@isegi.unl.pt Abstract The method proposed in this paper supports the UAV network path definition in an autonomously way, taking into consideration the density of the detected events at each moment, in each place. We use the selforganizing maps to detect event patterns in the field of view of the sensors, allowing unmanned aerial vehicles (UAV) path definition based on events. The goal of this paper is to maximize the detection of ships using a UAV network.. Introduction Technological advances in communication systems and the ability to build low-power and inexpensive mobile systems make the deployment of a group of networked vehicles a feasible task. Recent development of unmanned aerial vehicles (UAV) and recent implementations have proven their benefits in several applications such as fire detection [], agricultural monitoring [2], ocean surveillance [3], traffic surveillance [4] and military [5, 6]. UAV capabilities are improving, and new methods for deployment and management are required to coordinate the operation of UAV networks. Controlling a network of UAV implies a set of definitions such as goal assignment, resource allocation and trajectory optimization problems [7]. Different approaches have been used in the past to solve UAV path planning problems and it is not our aim to give a survey of all of these. In [8], the authors use a Voronoi approach to define the UAV path. This Voronoi diagram is built based on predefined way points and each edge of the Voronoi diagram is assigned two costs: a threat cost and a length cost. This method deals with the problem of simultaneous arrival of multiple UAV at their targets, using threatavoiding trajectories. In [9] a dynamic programming method is presented to generate near-optimal trajectories for multiple UAV to cooperatively search for targets. A coordination framework in which the UAV network seeks out local maxima or minima in the environmental field is presented in []. The network has adaptation capabilities in response to the sensed environment in order to optimize its gradient climb. In [] a decentralized trajectory planning of multiple UAV is presented, using a receding horizon strategy based on mixed integer linear programming. The objective of this paper is to present a new method for a UAV network path definition in an autonomously way. The path definition is obtained in an iterative process, being defined for each instant the behavior each UAV should assume. The method presented allows the self organization of a network of UAV using self-organizing maps (), allowing realtime adjustment of the network. As study case we use a UAV network to monitor ships. This paper is organized as follows. In section 2 we present the proposed method. In section 3 the environment simulator is presented, being some simulation and tests present in the section 4. Finally, section 5 is devoted to the discussion of the main results obtained in this research work. 2. Proposed method The method proposed in this paper supports the UAV network path definition in an autonomously way, taking into consideration the density of the detected events at each moment, in each place. The goal is to control the UAV network so that the number of detected events in the field of view of the sensors, at each instant, is optimum. To guarantee this optimum detection we assume each UAV does not share their field of view with others.

2 Since our method is based in the detection of events density, we apply a well know neural network called Self-Organizing Map () [2] to detect these patterns. The basic idea of the [2] is to map the data patterns onto a n-dimensional grid of neurons or units. That grid forms what is known as the output space, as opposed to the input space where the data patterns are. This mapping tries to preserve topological relations, i.e., patterns that are close in the input space will be mapped to units that are close in the output space, and vice-versa. So as to allow an easy visualization, the output space is usually or 2 dimensional. In our application, we use a 2-dimensional. Each unit represents an UAV, and the weights of the unit represent the UAV coordinates in the study area. Each ship is a data point, characterized by its two coordinates and the is repeatedly being trained, using the position of the detected ships as input data points. Each training phase will produce a new selforganized map, and each unit s weights will correspond to the new UAV coordinates. However, there are some UAV physical restrictions that need to be considered in the training phase. The has to be aware of the UAV maximum speed and to the fact that their fields of view must not intersect. Due to this fact the standard algorithm [3] was changed to fulfill these two restrictions. For experimental purposes we will assume a UAV network that is deployed to perform ship detection on a specific area of interest. The UAV network is composed by several UAV equipped with a camera and a radio transceiver. It is not our purpose to study the detection process or the communication issues since extensive work has been made in this field [4, 4]. 2.. The UAV path definition algorithm Randomly define each ship position (ShipsPos) 2 Repeat 3 Simulate the ships behavior (movement) in the area of interest for the instant t (update ShipsPos) 3 Define the detected ships based on the UAVpos, ShipsPos and UavSen 4 Train the according to the ships and UAV positions, taking in consideration the UavVel and UavSen 4 Move the UAV according to the units weights (update UAVpos) 6 Increase t 7 Until t reaches t f 3. The environment simulator To simulate the ships and its paths a ship simulator was implemented in Matlab, using a set of predefined characteristics such as type of ship (fishing or merchant), the number of ships and the minimum and maximum velocities for each type of ship. Each ship is initially randomly created, being the total number of ships present in the area of interest constant. The type of ship defines its behavior, i.e. a fishing boat will stay in the same area for a while, moving afterwards to a different place while a merchant ship has a constant direction and velocity (Fig. ). a) The UAV management algorithm based on may be described as follows: Let tf be simulation duration x,y,x,y define the area of interest UavVel be the maximum UAV velocity UavSen be the maximum UAV sensing range UAVpos be a collection of coordinates x i,y i defining the UAV initial positions ShipsPos be a collection of coordinates x i,y i defining the ships initial positions b) Fig.. The ship simulator. a) fishing boats versus merchant ships behavior b) initial distribution of the ships

3 4. Experimental evaluation To measure the effectiveness of the proposed method, we must define a metric by which it can be assessed, and some benchmark methods with which we may compare it. Since the objective is to monitor as best as possible the points of interest, an obvious metric is the number of ships observed, by the UAV, for the instant t, divided by the total number of ships present in the area of interest for the same instant. We call this ratio the instant coverage level (icl). t icl t = () TotalShipst a) sensor zigzag sensor Another metric we use is the global coverage level, defined as the percentage of the total detected ships until the present time (gcl). untilnow gcl present = (2) TotalShipsuntilNow Finally we calculate the average time of ships monitoring (amt). This metric will evaluate if a UAV can follow a moving ship. N = MonitoringTime i i amt = (3) where the MonitoringTime i is the amount of time ship i is being observed and is the total number of detected ships. 4.. The benchmark UAV algorithms As for a comparison benchmark, we chose two different methods for defining patrol itineraries. In the first method we assume fixed sensors in a regular matrix along the area of interest. In the second method we define each UAV a region to cover in the area of interest, and each one will follow a zigzag trajectory covering its correspondent region. Fig. 2a exemplifies the benchmark methods behavior in the area of interest. In Fig. 2b a 3x3 UAV network is presented along with the two different type of ships. b) Fig. 2. Benchmark methods: a) fixed and zigzag trajectories sensors b) based UAV network. 5. Results In Fig. 3 we present the instant coverage level, the global coverage and the average monitoring time for the base and the two benchmark methods. In this case, and due to the random ships' position, we calculate an average of simulations..4.2 Percentage of sensed Ships Number of diferent ships sensed/total ships Average Time on sensed ships Fig. 3. Statistics for and benchmark methods: fixed and zigzag trajectories sensors. From this figure we can conclude that our method presents the best instant covering, while the fixed and

4 zigzag trajectories methods presents similar instant coverage. However, the zigzag method presents a higher global coverage, detecting almost all ships that cross the area of interest. Our method only captures approximately 5% of the crossing ships. Inversely, in the average monitoring time the method presents a higher value, representing a higher capacity on tracking the ships once they are detected. In order to better understand the method we evaluate the same metrics changing the number of sensors and the number of ships. 5.. Changing the UAV number In this test, eight different UAV networks were used (with 4, 9, 6, 25, 36, 49, 64 and 8 UAV) and its instant coverage level calculated (Fig. 4). From the analysis of Fig. 4 we conclude, as expected, that an increase in the number of UAV imply a higher instant coverage. Fig. 4. Instant coverage level calculated for 8 sets of UAV Changing the ships number For this test we used a set of 9 UAV and we increased the number of ships present in the area of interest (Fig. 5). From the analysis of this figure we conclude, as expected, that an increase in the number of ships to detect by the UAV network, results in a smaller coverage of the area. Fig. 5. Instant coverage level calculated using 9 UAV and increasing the number of ships in the area of interest 6. Discussion In this paper we presented a first approach to the use of Self-Organizing Maps as a UAV network path definition model. The working example use in this study was based on ships detection in the ocean. The tests made prove that the proposed method improves the number of ships instantaneously detected using the based method, when compared to the benchmark methods. As future work we would like to include aerodynamics restrictions of each UAV, such as the curving angles, maximum and minimum acceleration and breaking. 7. References [] L. Merino, F. Caballero, J. R. Martinez-de Dios, and A. Ollero, "Cooperative Fire Detection using Unmanned Aerial Vehicles," presented at Robotics and Automation, 25. ICRA 25. Proceedings of the 25 IEEE International Conference, 25. [2] S. R. Herwitz, S. Dunagan, D. Sullivan, R. Higgins, L. Johnson, Z. Jian, R. Slye, J. Brass, J. Leung, B. Gallmeyer, and M. Aoyagi, "Solar-powered UAV mission for agricultural decision support," presented at Geoscience and Remote Sensing Symposium, 23. IGARSS '3. Proceedings [3] J. C. Rubio, J. Vagners, and R. Rysdyk, "Adaptive Path Planning for Autonomous UAV Oceanic Search Missions," presented at Proceedings of the AIAA st Intelligent Systems Technical Conference, 24. [4] K. Kaaniche, B. Champion, C. Pegard, and P. Vasseur, "A Vision Algorithm for Dynamic Detection of Moving Vehicles with a UAV," presented at Robotics and Automation, 25. ICRA 25. Proceedings of the 25 IEEE International Conference 25. [5] D. J. Nowak, I. Price, and G. B. Lamont, "Self organized UAV swarm planning optimization for search and destroy using SWARMFARE simulation," in Proceedings of the 39th conference on Winter simulation: 4 years! The best is yet to come. Washington D.C.: IEEE Press, 27. [6] U. Zengin and A. Dogan, "Real-Time Target Tracking for Autonomous UAVs in Adversarial Environments: A Gradient Search Algorithm," presented at Decision and Control, 26 45th IEEE Conference on, 26. [7] M. Alighanbari, Y. Kuwata, and J. P. How, "Coordination and control of multiple UAVs with timing constraints and loitering," presented at American Control Conference, 23. Proceedings of the 23, 23. [8] R. W. Beard, T. W. McLain, M. A. Goodrich, and E. P. Anderson, "Coordinated target assignment and intercept

5 for unmanned air vehicles," Robotics and Automation, IEEE Transactions on, vol. 8, pp , 22. [9] M. Flint, M. Polycarpou, and E. Fernandez-Gaucherand, "Cooperative control for multiple autonomous UAV's searching for targets," presented at Decision and Control, 22, Proceedings of the 4st IEEE Conference on, 22. [] P. Ogren, E. Fiorelli, and N. E. Leonard, "Cooperative control of mobile sensor networks:adaptive gradient climbing in a distributed environment," Automatic Control, IEEE Transactions on, vol. 49, pp , 24. [] T. Schouwenaars, "Safe Trajectory Planning of Autonomous Vehicles," in Department of Aeronautics and Astronautics, vol. Doctor of Philosophy: Massachusetts Institute of Technology, 26. [2] T. Kohonen, Self-Organizing Maps, 3rd edition ed. Berlin: Springer, 2. [3] J. Vesanto, J. Himberg, E. Alhoniemi, and Parhankangas, "Self-organizing map in Matlab: the Toolbox.," presented at Proceedings of the Matlab DSP Conference, Espoo, Finland, 999. [4] S. Todorovic and M. C. Nechyba, "Intelligent missions for MAVs: visual contexts for control, tracking and recognition," presented at Robotics and Automation, 24. Proceedings. ICRA '4, 24.