Reinforcement learning based hyperparameter optimization for irregular shape extended object tracking

Abstract The great leap forward in the development of artificial intelligence technology has brought a new prospect to optimal filtering technology. Extended object tracking (EOT), a well-known research hotspot, is optimized using reinforcement learning (RL) in this paper. Generally, in estimating star-convex extended objects using Gaussian process (GP), the selection of hyperparameters significantly impacts the estimation outcomes. Typically, a priori or predetermined set of fixed values constitutes these hyperparameters before the GP of learning commence. However, improper choices in hyperparameter selection may lead to suboptimal performance in practical shape estimation. Addressing this concern, this study introduces a hyperparameter optimization approach for GP-based EOT. This approach incorporates RL via Q-learning (QL) into GP of the extended object model, guiding the agent towards an adaptive hyperparameter strategy through a well-designed reward function. Consequently, this method effectively optimizes the selection of hyperparameters, progressively improving the performance of the multi-feature estimation, such as object trajectory and shape in EOT. Simulation experiments conducted across multiple tracking scenarios demonstrate that the proposed algorithm outperforms traditional GP and random hypersurface algorithms in estimation precision.