Adaptive elliptical extended object tracking based on deep reinforcement learning

Abstract To overcome the challenges arising from the estimation of elliptical shape parameters (orientation and semi-axis length) and unknown multiplicative noise covariance (MNC) in extended object tracking (EOT), this study proposes an innovative adaptive EOT method based on the soft actor-critic deep reinforcement learning (DRL) algorithm. Multiplicative noise, as a complex uncertainty factor, significantly affects the tracking performance of extended objects (EO), especially the recognition of shapes. Improper processing of MNC may lead to unstable parameter estimation or significant bias. The remarkable development of artificial intelligence, especially DRL, has provided a strong drive for intelligent decision making of optimal estimation. Through the combination of DRL and recursive Kalman filter, the estimation accuracy of EO orientation and elliptic axis length can be jointly improved, thus optimizing EOT performance to a large extent. This approach enhances the adaptability of the EOT algorithm through interactive learning between the agent and the environment with online adaptive tuning of the MNC and elliptical parameters. The simulation results show the superiority of the method compared with the existing methods, with significant improvement in estimation accuracy and robustness. In various test scenarios, the method can estimate the state and shape parameters of the EO more accurately, in which, for the shape estimation accuracy of the EO, the Gaussian Wasserstein distance of the proposed algorithm is optimized by 20.3%, 23.7%, and 18.9%, respectively, compared with the existing optimal method in different experimental scenarios.