Intelligent IR/4K Collection for Enhancing Long Range Aerial Detection

The rapid advancement of autonomous drone traversal creates an urgent need for high-quality, diverse training data. Machine learning models for critical tasks like navigation and object detection depend entirely on the datasets used to train them. While the existing Long-Range Drone Dataset (LRDD) from the iMaple lab provides substantial volume, significant gaps remain, particularly in infrared imagery, 4K high resolution and balanced environmental representation. Current drone detection datasets, including LRDD, suffer from core deficiencies that limit model performance, namely a lack of rigorous classification for image attributes (e.g., lighting, altitude, background), and identifiable biases where certain scenarios are overrepresented while others are critically underrepresented. The imbalances in image diversity create ambiguous relationships between dataset composition and model improvement. The project therefore aims to systematically improve the iMaple dataset by establishing a data collection protocol that verifiably enhances dataset quality through targeted diversity, and by directly contributing a large volume of strategically chosen, labeled images to correct existing gaps.
Our approach is methodical and data-driven, combining targeted field collection with analytical validation. First, we will develop a refined image attribute classification system—extending beyond standard categories—to guide collection. We will then execute synchronized 4K RGB and thermal infrared drone flights under a wide variety of controlled environmental conditions (altitude, lighting, weather) to capture the needed diversity. This new data will be processed, labeled, and integrated into LRDD. By training baseline detection and evaluating performance gains as new data is added, we will develop a metric to distinguish between improvement from sheer volume versus improvement from attribute diversity. This process ensures our contributions are not only additive but demonstrably enhance the dataset's overall quality and utility for building robust machine learning models.