A Data-Driven Maneuvering Target Tracking Method Aided with Partial Models

Target tracking plays a vital role in both civil and military fields. Traditional radar point (maneuvering) target tracking methods always require a prior kinematic model to match the target motion. In fact, it is hard to satisfy this requirement in practice, especially when sudden maneuvering happens for non-cooperative target tracking, which leads to an undesirable tracking peak error. Motivated by this, a new data-driven maneuvering target tracking method aided with partial models is proposed in this article to suppress this peak error. Here, since the historical target tracks include as many kinds of sudden maneuvers as possible, a data-driven learning-based network is trained to obtain high-precision state estimate when the target makes unpredictable maneuvers. Meanwhile, when the target has weak or no sudden maneuver, state estimate with satisfying accuracy is still obtained via prior kinematic models, where the Kalman filtering gain and model parameters are learned through the designed network to further promote its adaptivity. In particular, the kinematic modeling-based estimation also maintains the tracking robustness in the absence of representative training data. In addition, a discriminant network of evolution models is constructed to decide the dominant model (i.e. data-driven evolution model or one of the kinematic models) online based on radar measurements in the sliding window, which actually outputs the weights to make the final estimate as a weighted combination of the data-driven learning-based estimate and the kinematic modeling-based estimate. The proposed method is discussed in detail from the perspectives of tracking precision comparison, timeliness analysis, and noise analysis. The results show that the algorithm can ensure the timeliness of calculation and has a certain robustness under different noises. And most importantly, it has higher tracking precision when target maneuver happens suddenly.