A Novel Track Segment Association Method Using Contextual Contrasting

During target tracking, interruptions often occur due to target maneuvers, missed detections, and long sampling intervals. These interruptions can significantly hinder information fusion and situational awareness. Traditional methods, which rely on hypothetical models and estimation theories to measure similarity between track segments to be associated, have several drawbacks such as unreasonable hypotheses, unsuitable models, and uncertain thresholds. To address these issues, data-driven methods with strong learning ability, like deep learning methods, are gradually used. However, existing data-driven methods are difficult to generalize to scenarios with long interruption interval condition. In order to tackle this problem, a novel track segment association method using contextual contrasting (TSA-CC) is proposed. TSA-CC consists of three key components: (i) separation of track segment into subseries-level patches, which serve as input tokens for the Siamese Transformer encoders; (ii) introduction of learnable time-related embeddings to distinguish temporal distance between sampled segments; (iii) contextual contrasting module used for learning discriminative representations. TSA-CC then selects the track segments corresponding to the nearest vectors in the feature space as associated pairs. Experiments on a real-world AIS dataset demonstrate the effectiveness of TSA-CC.