Two-Stage Spatio-Temporal Feature Correlation Network for Infrared Ground Target Tracking

Similar target distractor and background occlusion in the complex ground environment can result in infrared target tracking drift or even failure. To solve this problem, this study proposes an infrared ground target tracker based on a two-stage spatio-temporal feature correlation network. First, a spatio-temporal context fusion feature correlation network (Scffcnet) is proposed, which fuses appearance features and spatio-temporal context information, and improves the stable tracking ability of the tracker under similar target distractor conditions. Second, a unidirectional trajectory feature correlation network (UTFCNet) is proposed, which ensures the accurate prediction of ground target trajectories by effectively using the temporal context information and optimizing training and application methods. Finally, a two-stage anti-occlusion strategy of 'occlusion-prediction-recapture' is proposed, which improves the anti-long-term occlusion performance of the tracker. Qualitative and quantitative experiments on image sequences under similar target distractor and background occlusion conditions verify the effectiveness of the proposed tracker.