A real-time drone detection framework for variable environmental conditions based on web-based deep learning

Abstract

The rapid proliferation of unmanned aerial vehicles (UAVs) in civilian, industrial, and military domains has intensified challenges related to airspace security, particularly in restricted and sensitive areas. Unauthorized drone operations pose risks including illegal surveillance, disruption of air traffic, and threats to critical infrastructure. This research proposes a lightweight, web-based real-time drone detection framework utilizing a convolutional neural network (CNN) integrated with a fast application programming interface (FastAPI) server to support continuous visual surveillance. The system is trained and evaluated using a custom dataset of 23,118 annotated drone images, enhanced through data augmentation to improve robustness under varying environmental conditions. Experimental results demonstrate real-time inference performance of 20–25 frames per second across multiple camera resolutions, including 1080p, 2K, and digital single-lens reflex (DSLR) Canon 600D. Detection accuracy varies between 53% and 74%, revealing a strong dependency on illumination and imaging conditions, with optimal performance achieved under high illumination levels (45,000 lux). Performance degradation is observed in low-light environments, particularly for DSLR-based detection. The proposed system enables live video streaming and static image analysis through a user-friendly web interface without reliance on external services. Overall, this study presents a practical and deployable prototype for automated drone monitoring, emphasizing the trade-offs between accuracy, illumination, and camera resolution in real-world airspace security applications.