Data-driven online tracking filter architecture: A LightGBM implementation

Abstract

State estimation has been dominated by traditional algorithms based on the Bayesian framework, and Bayesian filtering has proven to be an effective component with a wide range of applications in various fields. On the other hand, Bayesian framework is inherently limited by the precise model prior information, while struggling to cope with the challenges such as Markov assumptions that do not adequately describe the actual motion modes, and model transition decision delay in the multiple model algorithm. To alleviate this, we redefine the state filtering problem from a data-driven perspective as a non-probabilistic mapping problem from measurement sequence to target state, replacing the iteration of model derivation with supervised learning regression. Specifically, we introduce a data-driven tracking filter architecture (DTF) that includes spatio-temporal feature processing and decoupling of target trend and sensor uncertainty features. Its fast implementation based on LightGBM is further given, considering the interpretability of the model. Simulation results show that the proposed method approaches to the theoretical optimal solution of the Kalman filter in an ideal linear Gaussian scenario. In challenging maneuvering target tracking scenarios, the proposed model is superior in performance compared with other existing methods. Finally, we validate the effectiveness of the method in radar tracking and real-world video scenes.