Bird and drone detection and tracking is a critical problem in air security. These targets typically appear far away and small, making it difficult to obtain sufficient visual cues for identification. The latest detectors use multi-scale feature pyramids to handle each target size, but still fail when dealing with extremely small objects.Deep features of the lower layers emphasize semantic information, whereas higher-resolution features of the upper layers capture spatial information useful for location search; designing a well-structured deep neural network that complements these two feature types is a central issue in object detection. This paper proposes anovel model based on YOLO26-P2 [1—3] for detecting micro aerial targets. Using the detection capability of the higher-resolution P2/4 feature as its base, we propose a P3-to-P2 static fusion model and evaluate whether it improves the P2 feature. Next, we propose a modified P3-to-P2 semantic feedback model and evaluate its performance against the static fusion model. Unlike the static fusion model, the proposed method reinserts the modified P3 feature into the P2 layer, represented as P2 → Modified P3 → Modified P2. We set four models—the standard model [4, 5], the high-resolution P2 model, the P3-to-P2 static fusion model andthe P3-to-P2 semantic feedback model—and compare each model to evaluate the performance of theproposed models. Finally, the ByteTrack [21] tracking algorithm was combined with the final detector to implement MOT (multi-object tracking). The proposed method introduces simple changes to the feature layers within YOLO26 without any external image processing. In this paper, we set up a testing plan that considers both the computational load and real-time performance to evaluate the performance of the proposed method.