TinyML-On-The-Fly: Real-Time Low-Power and Low-Cost MCU-Embedded On-Device Computer Vision for Aerial Image Classification

Aerial image classification is essential to intelligent surveillance and monitoring systems. Traditional computer vision methods either uses computational offloading to high-end servers or edge devices. However, unmanned aerial vehicles (UAVs) platforms have resource and power constraints. Aerial image classification is complicated and less-expensive UAVs lack processing power and cameras. Even with large-scale computing environments, methods for classifying images are difficult to apply to aerial imagery. We propose TinyAerialNet leveraging TinyML for real-time inference on a resource-constrained ESP32 CAM. The model tested on AIDER dataset, achieves 88% on-device accuracy in the micro-controller with 103.9 KB RAM and 850 milliseconds for inference.