Improvement of on-board object identification performance of autonomous nano-drones through the use of deep learning techniques and neural architecture search.
Parole chiave APPRENDIMENTO PROFONDO, BASSO CONSUMO, DRONI, EFFICIENZA ENERGETICA, INTELLIGENZA ARTIFICIALE, MICROCONTROLLORI, RETI NEURALI CONVOLUZIONALI, RETI NEURALI PROFONDE, ROBOTICA, SISTEMI EMBEDDED, UAV
Riferimenti DANIELE JAHIER PAGLIARI
Riferimenti esterni Dr. Alessio Burrello (Politecnico di Torino), Dr. Lorenzo Lamberti (University of Bologna), Dr. Daniele Palossi (SUPSI-IDSIA, Lugano)
Tipo tesi SPERIMENTALE, SVILUPPO SW
Descrizione The candidate will focus on enhancing the performance of autonomous nano-drones in environmental exploration and object identification tasks, by using deep learning techniques and neural architecture search (NAS) algorithms. The scope of the study revolves around the creation of synthetic data from a simulator to improve the end-to-end performance of the drone. Specifically, during this thesis, the candidate will try to reach the following milestones:
1. Propose a new NAS algorithm that is specifically designed to enhance the object recognition performance of autonomous nano-drones.
2. Include data generated by a simulator in the NAS search loop, to achieve an improvement of the end-to-end environment exploration and object identification task performed by the nano-drone.
3. Test the final neural network found by the NAS in-field on real-world data.
The study aims to achieve better end-to-end performance by, co-optimizing the object recognition performance on individual frames, by reducing false positives and negatives, and the processing time required by the neural network, which is crucial for the efficient operation of autonomous nano-drones.
Vedi anche screenshot from 2022-04-21 20-30-37.png
Conoscenze richieste Required skills include C and Python programming. Further, a basic knowledge of computer architectures and embedded systems is necessary. Lastly, the candidate should also be familiar with machine/deep learning algorithms and models.
Note Thesis in collaboration with Prof. Luca Benini and Dr. Lorenzo Lamberti, from the University of Bologna, and Dr. Daniele Palossi from SUPSI – IDSIA, Lugano.
Scadenza validita proposta 06/06/2022 PROPONI LA TUA CANDIDATURA