An Optimization Framework for Efficient Vision-Based Autonomous Drone Navigation

Fully autonomous drones are a new emerging field that has enabled many applications such as gas source leakage localization, wild-fire detection, smart agriculture, and search and rescue missions in unknown limited communication and GPS denied environments. Artificial intelligence and deep Neural Networks (NN) have enabled applications such as visual perception and navigation which can be deployed to make drones smarter and more efficient. However, deploying such techniques on tiny drones is extremely challenging due to the limited computational resources and power envelope of edge devices. To achieve this goal, this paper proposes an efficient end-to-end optimization method for deploying deep NN models for vision-based autonomous drone navigation applications, such as obstacle avoidance and steering task. This paper formulates two different methods for implementing the NN inference phase onto tiny drones and analyzing the implementation results for each case: 1) a Cloud-IoT implementation and 2) Onboard Processing. Several models are trained with state-of-the-art scalable NN architectures and the most efficient cases in terms of computation complexity and accuracy are selected for implementation on a cloud server and several edge devices. By designing hardware-friendly NN models and optimal configuration of the implementation platforms, we were able to reach up to 97% accuracy, speed up the computation 2.3×, have 22× less complexity, and 53% energy reduction. Also, we achieve up to 25 fps on the GAP8 processor, which is enough for real-time drone navigation requirements, even when the model is running on a small IoT device.