Adaptive Deep Neural Networks for Human Pose Estimation on Autonomous Nano-Drones

Fully autonomous nano-scale unmanned aerial vehicles (UAVs), due to their tiny form factor and compact size, are particularly suitable for indoor applications that require safe navigation close to humans. These applications encompass surveillance, monitoring, ambient awareness, and interaction with smart environments. Recent research shows that deploying AI models in Nano-Drone systems with onboard computations offers several advantages, including reduced operational costs, minimized inference latency for real-time scenarios, and enhanced data security and privacy. However, these small devices face severe constraints in terms of memory and computational resources. Adaptive or Dynamic Neural Networks are an emerging research topic in deep learning. These Networks can adapt their structures or parameters based on the input during inference, thus efficiently allocating computations on demand, selectively activating model components (e.g., layers, channels, or sub-networks). The objective of this thesis is to apply adaptive inference techniques to construct networks for estimating and maintaining the relative 3D pose of a nano-UAV with respect to a moving person in the environment. The task requires mapping a low-resolution image to the relative pose of the subject, consisting of 3D coordinates (X, Y, Z), and a rotation angle (ϕ) with respect to the gravity Z-axis. We propose an input-dependent adaptive inference methodology based on the Big/Little model approach. This approach is based on constructing two models, a little, efficient, but less accurate model, and a bigger and more accurate model, executed in sequence with an early-stopping decision function with the goal of reducing the executions of the big model. By utilizing pre-trained models optimized for the task, our methodology aims to reduce computational cost and inference latency with minimal changes in network regression performance. We constructed three networks to test each methodology, fixing one big model and alternating the Little model. The proposed approach has yielded promising results, achieving a significant reduction (>30%) in inference cycles at iso-accuracy, with respect to static models.