Real time UAV autonomy through offline calculations

Sunghun Jung, Purdue University

Abstract

Two or three dimensional mission plans for a single or a group of hover or fixed wing UAVs are generated. The mission plans can largely be separated into seven main parts. Firstly, the Region Growing algorithm is used to generate a map from 2D or 3D images. Secondly, the map is analyzed to separate each blocks using vertices of blocks and seven filtering steps. Thirdly, the Trapezoidal map algorithm is used to convert the map into a traversability graph. Fourthly, this process also filters out paths that are not traversable. That is, nodes located inside the blocks and too closely located nodes are filtered out. Fifthly, the Dijkstra algorithm is used to calculate the shortest path from a starting point to a goal point. Sixthly, the 1D Optimal Control algorithm is applied to manipulate the velocity and acceleration of the UAVs efficiently. Basically, the UAVs accelerates at one graph node and maintains a constant velocity and decelerates before reaching the next graph node. Lastly, Traveling Salesman Problem Method (TSP) algorithm is used to calculate the shortest path to search the whole region. After this discretization of space and time, it becomes possible to solve several autonomous mission planning problems. We focus on one of the most difficult problems: coordinated search. This is a multiple Traveling Salesman Problem (mTSP). We solve it by decomposing the search region and solving TSPs for each vehicle searching a sub-region. The mTSP is generally used when there are more than one salesman is used. In addition to the four main parts, there are three minor parts which support the main parts. Firstly, Target Detection algorithm is generated to detect a target located near the UAVs' path. A picture of the desired target is inserted into the algorithm before UAVs launch. Using the Scale-Invariant Transform Feature (SIFT) algorithm, a target with a specific shape can be detected. Secondly, Tracking algorithm is generated to manipulate UAVs to follow targets. Once one or several targets are detected, UAVs near the target are manipulated to approach to the target. If the number of detected targets is more than one, UAVs are evenly grouped to track targets. After a specific period of time, UAVs hand off and continue their original tasks. Thirdly, Emergency algorithm is generated to avoid losses of UAVs when UAVs have system failures. If one UAV is out of fuel or control during the mission, the Emergency algorithm brings the malfunctioning UAV to the point of departure and let the rest UAVs to continue an aerial reconnaissance. An UAV which finishes its task the earliest will continue to search a region which the failed UAV is supposed to search. In addition, Emergency algorithm prevents UAVs colliding into each other by using emergency altitude. Overall, the framework developed here facilitates the solution of several mission planning problems. The robustness built into our discretization of space and time permits feedback corrections on real-time to vehicle trajectories. The library of off-line solutions proposed and developed here minimizes computational overhead during operations.

Degree

M.S.M.E.

Advisors

Ariyur, Purdue University.

Subject Area

Aerospace engineering|Mechanical engineering|Computer science

COinS