6-D pose tracking within a quadplane swarm using particle filter with KAPAO network and 3D-error enhancement

Abstract

Accurately measuring the relative 6-D pose between unmanned aerial vehicles (UAVs) within a formation is fundamental for UAV swarms to execute tasks effectively. Existing monocular 6-D pose estimation and tracking methods struggle with pose ambiguity when UAVs are widely spaced. This paper proposes an improved particle filtering method for quadplane 6-D pose tracking to eliminate ambiguity and enhance accuracy. Our method integrates a KAPAO network as an observation model to handle complex image backgrounds, combined with a constant velocity motion model to adapt to the diverse motion states of quadplanes. We utilize the 3D object-space collinearity errors for weight updating to enhance adaptability to the images captured by an airborne zoom camera and fully leverage quadplane motion information to prevent algorithm divergence. Both point and line errors in updating the weights for position and orientation separately help mitigate their mutual coupling effects, ultimately enhancing overall accuracy. Our approach performs exceptionally well on quadplane datasets by eliminating pose ambiguity and maintaining the upper bounds and medians of the 3D error box plots respectively below 3.19 and 0.96 meter for distances ranging from 31.6 to 100.0 meters between two quadplanes. Furthermore, the ADD and Rete accuracy indicators are also $1-3$ times higher than some top-tier methods, with a runtime of just 35.2 milliseconds. This positions it as a promising solution for practical air-to-air quadplane missions.