Implicit Optical Flow-Guided Network for Infrared Small-Target Tracking

Infrared small-target tracking (ISTT) holds significant practical value in both military and civilian applications. Currently, some studies achieve ISTT by using optical flow-based methods. However, existing optical flow methods typically rely on pixel-level motion estimation or use flow maps as auxiliary network input and struggle to apply flow fields directly for bounding box localization, which further degrades tracking stability. Furthermore, in low signal-to-noise ratio infrared sequences, explicit optical flow frequently exhibits high errors and costs. To address these challenges, we propose a complementary and optical flow-like fusion network (COFNet), which constructs a flow-like field at high-level feature maps to produce temporally enhanced feature maps that improve cross-frame consistency of bounding boxes and enable robust association. Specifically, we propose an optical flow-like feature fusion (OFF) module that computes interframe similarity and displacement relationships across multiscale feature maps and temporally corrects current features while integrating cross-scale information, thereby making the generated bounding boxes easier to associate. This module establishes an implicit flow field at the feature level to utilize interframe motion information without the computational cost of explicit optical flow, thereby improving robustness to occlusion and localization stability during tracking. To compensate for lost positional cues and reduce mismatches caused by box displacement during tracking, we design a channel splitting and complementary mapping (CSM) module that divides feature channels into semantic and spatial subspaces and fuses them via recalibration. To suppress false positives and missed detections that disrupt trajectory continuity, a block-wise self-attention (BSA) module that enhances responses to fine textures and local context around the target. Finally, we employ DeepOC-SORT at the detection output to identity assignment. Experimental results on public datasets demonstrate that the proposed method attains superior tracking performance.