End-to-End Real-Time Drone-Based Person Detection Framework Using Deep Learning

This paper addresses the challenge of performing real-time person detection from Unmanned Aerial Vehicles (UAVs). Traditional computer vision methods often fail in aerial contexts due to the drone's constant motion, varying altitudes, and changing illumination. The authors propose an end-to-end pipeline that streams live video from a drone to a GPU-accelerated workstation, where a fine-tuned YOLOv8-nano model performs inference. The system is designed to maintain high frame rates while providing reliable detection across different flight altitudes and camera angles.

The system architecture consists of three main components: live feed acquisition, GPU-accelerated inference, and performance monitoring. The drone captures and transmits video wirelessly to a ground station, which decodes the stream and processes it using the YOLOv8-nano architecture. The model was trained on the VisDrone2019 dataset, with categories merged into a single 'Person' class to optimize for surveillance. The authors implemented a real-time monitoring mechanism that calculates processing latency and overlays the frame rate (FPS) directly onto the video feed to ensure synchronization between the drone's flight and the detection output.

Quantitative evaluation on the VisDrone2019 test set yielded a precision of 0.574, a recall of 0.410, and an mAP@0.5 of 0.448. Field experiments conducted at IIT Guwahati confirmed the system's practical viability, showing that the model achieves near-total detection reliability at altitudes between 16 and 25 meters. The system consistently maintained a frame rate above 41 FPS, peaking at 50 FPS, which is sufficient for real-time aerial monitoring applications. Qualitative analysis showed that the model is robust to lighting variations but struggles with complex backgrounds containing objects that mimic human shapes.