Lightweight AI Vision Module for Obstacle Detection and Risk Estimation on Crazyflie Micro UAVs

Micro unmanned aerial vehicles (UAVs) require efficient and reliable perception systems to operate safely in complex environments. However, deploying deep learning-based vision algorithms on micro-UAV platforms remains challenging due to limited computational resources, memory capacity, and power constraints. This paper presents a lightweight vision-based perception framework for obstacle detection and risk estimation on a Crazyflie micro-UAV platform. A PyBullet-based simulation environment is developed to generate synthetic RGB camera observations with automatically generated obstacle distance and risk annotations. Using this dataset, a lightweight multi-task convolutional neural network based on Mobile Net V۳-small is trained to simultaneously estimate obstacle distance and classify collision risk. The proposed model is compared with a ResNet۱۸ baseline in terms of prediction accuracy and computational efficiency. The MobileNetV۳-small model achieves a distance estimation mean absolute error of ۰.۱۰۴۵ m and a risk classification accuracy of ۹۵.۳۳% while requiring only ۱.۰۸ million parameters and achieving ۱۵۵.۷۸ FPS inference speed. Although ResNet ۱۸ provides slightly higher accuracy, it requires significantly greater computational resources, making it less suitable for micro-UAV applications. Furthermore, the proposed framework is validated using real images collected from a Crazyflie ۲.۱ platform equipped with a lightweight XIAO ESP۳۲ camera module. To address the simulation-to-real domain gap, a lightweight adaptation strategy using a small number of real images is introduced, reducing the real-world distance estimation error from ۰.۹۱۹۲ m to ۰.۲۰۶۶ m. The obtained results demonstrate the feasibility of lightweight AI-based perception for resource-constrained micro-UAV platforms.

کلیدواژه ها:

Micro UAV ، Obstacle detection ، Lightweight deep learning ، Computer vision