Particle filtering based optimization applied to 3D model-based estimation for UAV pose estimation

To estimate the pose of an unmanned aerial vehicle (UAV) during the landing process aboard a ship is used a vision system based on a standard RGB camera using one workstation for processing data. It is used a ground-based vision system to allow the use of a small size and weight UAV, due to the low computer requirements onboard. The resampling step in the particle filter takes a decisive role in the obtained pose estimation, and ten traditional resampling steps are tested and compared with a developed resampling strategy inspired by the genetic algorithms. The obtained results show that a classical resampling step is not sufficient in this problem and gets easily stuck in local minima. Those local minima are originated by the high dimensional search space and from the employed observation likelihood metric, which is very dependent on the UAV geometry which generates ambiguity in the pose estimate for complementary poses (poses with large wing pixel overlap, where the majority of the UAV area is concentrated). Results show that the errors obtained are lower and much more compatible with the requirements for this kind of problems when compared to the common existing resampling strategies.