Neural architecture search for real-time object detection in drone navigation

The surge in demand for autonomous flying drones has initiated research in real-time obstacle detection and avoidance. This paper considers the application of Neural Architecture Search (NAS) to enhance object-detection networks for drones with tight latency and power constraints. The proposed NAS approach balances accuracy, inference time, and energy consumption through multi-objective optimization. Tests with simpler datasets and actual tests demonstrate how NAS-optimized models bring inference down to 27.4 ms per frame, slice power consumption by 25%, and increase flight duration by 8 to 10% over YOLOv5 and SSD models. While the mean Average Precision (mAP) is slightly down at 65.3%, the NAS model achieves a 92% success rate for obstacle avoidance and keeps false detections below 3%. It surpasses the others in responsiveness and stability. These results point to NAS as a useful, software-only approach to improve real-time perception of UAVs and provide safe autonomous navigation in dense environments.