Smart Adaptive Perception for High-Precision Lightweight Infrared UAV Detection and Tracking

The unauthorized intrusion of unmanned aerial vehicles (UAVs) into restricted areas poses a significant challenge for effective detection and tracking, especially in low-visibility infrared environments. Existing methods struggle with real-time performance, complex backgrounds, and the tracking of small, distant, and blurry targets. To address these issues, we propose a novel lightweight framework for enhancing the detection and tracking capabilities of infrared UAVs. Our model integrates two biologically-inspired innovations: DySample, which performs adaptive upsampling to enhance the features of distant targets; and SimAM, a parameter-free attention module that enables robust background suppression and feature enhancement. Evaluation on a challenging customized infrared dataset demonstrates that our method achieves an mAP@0.5 of 85.0% (5.0% higher than the baseline) and uses only 7.03 million parameters. When combined with Bot-SORT, it achieves a tracking accuracy of 99.51% and an average center position error of 3.05, significantly reducing false alarms. This lightweight and high-performance solution is suitable for critical UAV monitoring.