Out-of-Distribution-Aware Contrastive Learning for Trajectory-Based Threat Identification

This paper presents a trajectory-based prototype-guided contrastive learning method for threat target classification and out-of-distribution(OOD) detection. To address softmax classifiers’ overconfidence and feature dispersion issues, our approach employs a GRU encoder to represent variable-length trajectory data and a projection head to transform these features for contrastive learning. Dual contrastive losses are applied at both the encoder and projection head levels to tighten intra-class representations and separate inter-class features. Furthermore, a nearest prototype classification scheme intrinsically detects OOD samples without external data. Experimental results on simulated radar trajectory datasets demonstrate that our method significantly outperforms conventional softmax-based models, especially in distinguishing challenging OOD samples with dynamics similar to in-distribution data.