Research on Tracking Algorithm of Infrared Security Small Target Based on YOLOv8

In response to the low efficiency and accuracy of infrared small target tracking algorithms in security scenarios, this paper proposes an improved algorithm that combines YOLOv8s with DeepSORT for infrared small target tracking. By introducing SSFF and TFE to reconstruct the Neck section, adding a secondary creative small target detection head, and designing a lightweight detection head, the algorithm integrates more scale information to enhance the recognition capability of small targets while maintaining detection speed. Experimental results show that the mAP of the improved SFsmall-YOLOv8s model in this paper reaches $93.7 \%$, with improvements of $1.3 \%$, $\mathbf{2 . 2 \%}$, and $2.3 \%$ in Precision, Recall, and mAP, respectively. Finally, the improved detector is fused with the tracker, and the REID feature extraction network is retrained. Testing shows that compared to the YOLOv8s+DeepSORT algorithm, the MOTA of the improved algorithm in this paper is increased by $4.56 \%$, adapting well to complex infrared human-vehicle small target tracking scenarios.