Integration of Deep-learning-powered Drone-to-drone Pose Estimation on Ultra-low-power Autonomous Flying Nano-drones
keywords ARTIFICIAL INTELLIGENCE, CONVOLUTIONAL NEURAL NETWORKS, DEEP LEARNING, DEEP NEURAL NETWORKS, DRONES, EMBEDDED SYSTEMS, ENERGY EFFICIENCY, FIRMWARE DEVELOPMENT, LOW POWER, MICROCONTROLLERS, ROBOTICS, SOFTWARE, UAV
Reference persons DANIELE JAHIER PAGLIARI
External reference persons Dr. Daniele Palossi (Università della Svizzera Italiana)
Research Groups DAUIN - GR-06 - ELECTRONIC DESIGN AUTOMATION - EDA, ELECTRONIC DESIGN AUTOMATION - EDA, GR-06 - ELECTRONIC DESIGN AUTOMATION - EDA
Thesis type EMBEDDED SOFTWARE DEVELOPMENT, EXPERIMENTAL, SOFTWARE DEVELOPMENT
Description Nano-sized unmanned aerial vehicles (UAVs) are key elements for human-drone interactions. For instance, the work that will be addressed in this work is fundamental for the estimation of the relative position in a drone’s swarm. In this scenario, the energy efficiency of the algorithms, together with high accuracy is essential for a successful drone mission. On the other hand, common powerful algorithms for computer vision are based on large and memory-hungry convolutional neural networks which can not be fitted on nano-drones with a power envelopment < 100mW. To tackle this gap, in this thesis the candidate will explore different solutions to maintain high accuracy in the task of drone-to-drone estimation while minimizing the dimension and the number of operations of the target algorithm employed.
In particular, the thesis will encompass two phases, the first one carried out al Politecnico of Torino and the second one at the University of Italian Switzerland, Lugano.
The first phase will include: i) Study of the literature in drone-to-drone pose estimation and in Neural Architecture Search; ii) Extension of already existing frameworks for NAS, to support more general networks. iii) Research of a Pareto set of architectures in the accuracy vs efficiency space iv) Deployment on the target platform, the multi-core MCU GAP8.
The second phase will include: i) The collection of a new dataset in the IDSIA facilities; ii) The extension of the NAS search to the new dataset, with the possible study of transfer-learning from the more general one; iii) Deployment of the new architectures on the nano-drone; iv) In-field test and benchmarking of the networks on autonomous nano-drones.
Notes Thesis in collaboration with Dalle Molle Institute for Artificial Intelligence (IDSIA) SUPSI-USI, Lugano, Switzerland. . The thesis will be carried out both in Torino and in IDSIA, SUPSI-USI East Campus, Lugano.
Deadline 24/09/2021 PROPONI LA TUA CANDIDATURA