Hierarchical Trustworthy Inference for UAV/Bird Classification Under Track Data Imprecision

Identifying targets as belonging to uncrewed aerial vehicles or birds exploiting radar track sequences is critical for flight operation management and airport critical infrastructure safety. In this context, a hierarchical classification framework with 'multifeature trustworthy integration (MFTI), multipreference decision fusion (MPDF), and multiwindow negotiation-based iterative reasoning (MWNIR)' is proposed. The MFTI employs a short-term multibranch deep evidential neural network, which leverages the Dirichlet process to reliably integrate the motion and radar cross section features of the target. The MPDF combines the outputs of two MFTIs with different preferences based on the Dempster-Shafer (DS) theory, providing reliable classification results on short-term track segments with different qualities. Finally, to reduce the impact of track data imprecision on the inference process, the MWNIR exploits a negotiation and iterative temporal DS-evidence fusion. Experiments on real-world radar track data demonstrate that the proposed algorithm can outperform some state-of-the-art methods in terms of classification accuracy and reliability.