Unsupervised and interpretable track‐to‐track association based on homography estimation of radar bias

Abstract

Track‐to‐track association methods based on machine learning and deep learning have greatly improved the association results, but the scope of application is limited by the poor interpretability and manual association labelling. To enhance the interpretability of the neural networks, enhance the credibility of association decisions, and reduce the consumption for labelling associated track pairs, the authors estimate and counteract radar bias by homography estimation to achieve track‐to‐track association. The proposed model is composed of a mixing extraction module and a homography estimation module. Aiming at the interaction of temporal and spatial features of tracks, the spatial‐temporal mixing features are extracted by a mixing extraction module. Focusing on attaining explainable discriminant factors, the homography matrix is generated by the homography estimation module. Targeting at the unsupervised learning, the radar bias and association matrix are estimated jointly so that the labelled track association pairs are not demanded. Finally, a track from one radar is transformed into the other radar, and the homography matrix that counteracts the radar bias can provide explainable discriminant factors and make the association decision more credible. Extensive experiments demonstrated that the proposed method can achieve better association results and the association results can be well interpreted.

Guest Editorial: Advances in AI‐assisted radar sensing applications

“…Xiong et al. tackle track‐to‐track association (T2TA) challenges by using homography estimation to address radar bias, enhancing association credibility and reducing manual labelling [5]. Perďoch et al.…”