Asymmetric IR Tracking with Cross-Scale Attention Interaction

The near infrared (NIR) target is significant in the recognition of UAV, whereas it is difficult to track small infrared targets under complex background. Due to the weak thermal feature, small size and noise like similar background, it is difficult to estimate state reliably. To this end, we proposed a novel lightweight dual-stream network for infrared target tracking. After the decoupling and feature matching, the proposed method can effectively represent and accurately localize small targets. A feature enrichment and multi-scale fusion mechanism is also proposed to enhance robustness, whereas structural optimization and parameter reduction guarantee a lightweight model. The experimental results on several infrared datasets show that our method has high tracking accuracy and real-time performance, which verifies the effectiveness of the proposed algorithm.