Deep Learning-based Soldier Detection in Forested Environments

Modern person detection models are typically trained on urban-centric datasets where targets are visually distinct from their surroundings. However, in forested environments, military uniforms and complex natural backgrounds significantly degrade detection performance. To address this gap, we collect a dataset of individuals wearing military uniforms in wooded areas, simulating realistic low-altitude search and surveillance scenarios. We demonstrate that state-of-the-art YOLO series detectors, widely used for real-time inference, show substantial performance drops when applied to these challenging settings. Using our dataset, we train and evaluate YOLO-based models to quantify their effectiveness in camouflage scenarios. Our findings highlight the limitations of existing detectors in recognizing individuals in military uniforms and provide a foundation for developing models tailored to such environments.