AIoT: a Kendryte K210 proof of concept

The Internet of Things is no longer a data source: in recent times, it is moving towards the deployment of Intelligent infrastructures thanks to the Artificial Intelligence. With the creation of new Systems-on-a-chip that able the Deep Learning on the edge, this technology has stopped being called I...

ver descrição completa

Detalhes bibliográficos
Autor: Giordano, Lucio
Tipo de documento: dissertação
Data de publicação:2020
País:España
Recursos:Universidad Complutense de Madrid (UCM)
Repositório:Docta Complutense
Idioma:inglês
OAI Identifier:oai:docta.ucm.es:20.500.14352/9045
Acesso em linha:https://hdl.handle.net/20.500.14352/9045
Access Level:Acceso aberto
Palavra-chave:004(043.3)
IoT
AIoT
Deep Learning
K210
Artificial Intelligence
Person detection
Yolo.
Informática (Informática)
1203.17 Informática
Descrição
Resumo:The Internet of Things is no longer a data source: in recent times, it is moving towards the deployment of Intelligent infrastructures thanks to the Artificial Intelligence. With the creation of new Systems-on-a-chip that able the Deep Learning on the edge, this technology has stopped being called IoT and it is being referred as AIoT: Artificial Intelligence of Things. One of these new Systems-on-a-chip that enable the Deep Learning on the edge, is the processor Kendryte K210. The Kendryte K210 is produced by Canaan and it contains a KPU, a neural network hardware accelerator, that allows the inference of deep neural networks on small and cheap commodity devices on which it is integrated. This work deals with the study and implementation of a neural network for person detection to be implemented on a Kendryte K210. Person detection is interesting because of its versatility of use: it may be used integrated on a security camera or to count the persons met in an environment. Also, it is useful being able to deploy a low-cost person detector without the need for powerful hardware or constant communications with the cloud. For that reason, the most advanced techniques techniques for object detection are analyzed, talking about the ones that achieve better results and are more suitable for the usage in a low-consumption co-processor. Then, we discuss the K210 capabilities, with an overview of its hardware and its software support, specifying how a network should be modelled to be run on the K210. We conclude stating that one possible approach for person detection on the K210, is to develop a Yolo V2 network using MobileNet as feature extractor and we discuss how the training has been done, what problems have been met during the process and how they have been solved.