Optimizing Deep Neural Network Architectures for Low-Power Autonomous Tiny UAV Navigation

Abstract

Interest in deploying Deep Neural Networks (DNNs) on the lowest-end embedded systems has grown over the past several years as DNNs have advanced in capability. However, as embedded systems are often computationally limited to fit a specific price point, energy consumption, and size, deploying a deep neural network without modification risks facing poor performance and high energy consumption due to it taking advantage of hardware accelerators simply not present on embedded systems, or improperly making use of the limited hardware accelerators that are present on the embedded system that is being targeted. In this project, we address the challenges of DNN implementation on resource-constrained devices and analyze the effect of different DNNs on power consumption, memory, and processor usage. Moreover, we successfully deploy DNNs on an expansion board for the Crazyflie series of tiny drones for onboard autonomous drone navigation while avoiding obstacles. We conducted a series of benchmarks on an embedded microcontroller specifically designed for running DNNs, comparing neural network architectures that, apart from a series of minor tweaks, were otherwise the same to uncover design factors that impacted the performance of the model when run on the hardware. After we changed the model by incorporating changes that alleviated poor performance, we found that we were able to net significant performance gains and reduce the overall system’s energy consumption.