Vision-based K-Nearest Neighbor Approach for Multiple Search and Landing with Energy Constraints

Abstract

The Flying Robot Trial League (FRTL) of RoboCup Brazil is a robotics challenge that aims to stimulate the study and development of autonomous and intelligent flying robots in the execution of different tasks applied to the industry. The challenge has four tasks. The first one aims to solve the problem of finding suitable landing places in an indoor environment. Landing often needs to occur suddenly and a suitable landing site needs to be found by the drone. This mapping is essential for an autonomous task to be performed safely. Thus, to solve this problem, unmanned aerial vehicles (UAVs) often needs to reconnaissance the area, and map the environment to detect mobile landing bases (that can be randomly allocated) and suspended landing bases. Thus, this work focuses on the use of UAVs for navigation in unknown GNSS-denied environments. This proposal combines the use of computer vision and K-nearest neighbor (KNN) to solve a travel salesman problem (TSP) in order to optimize the visitation of several landing pads by minimizing the flight time needed. The task benchmark is set to 10 minutes to be completed. However, our results demonstrate that our approach only takes 2.29 minutes to detect and visit all bases in this space.