Parallel Resampling for Accelerated Particle Filters in Vision-Based Terrain-Referenced Navigation

Particle filters require significant computational power due to the need to process a large number of particles for accurate state estimation. To meet real-time performance demands, parallelization of the particle filter, particularly a resampling process, is essential. In this paper, we propose a novel resampling algorithm tailored for parallel computing. The proposed method maximizes the independence of operations across particles while minimizing intercore communication, ensuring efficient use of graphics processing unit resources. A mechanism using a hard constraint is introduced to preserve particle distribution without excessive communication. The proposed method was tested using a simple Gaussian mixture model to assess how well the distribution is maintained. Additionally, the algorithm was tested in a vision-based terrain-referenced navigation system. Overall, the proposed resampling method demonstrated superior performance in terms of both accuracy and computational efficiency, mitigating the particle degeneracy problem and enabling real-time processing in computationally demanding environments.