Authors:
Graziele de Cássia Rodrigues
1
and
Ricardo Oliveira
2
Affiliations:
1
Computing and Systems Department, Universidade Federal de Ouro Preto, Joao Monlevade, Brazil
;
2
Computing Department, Universidade Federal de Ouro Preto, Ouro Preto, Brazil
Keyword(s):
Human Activity Recognition, Embedded Devices, TensorFlow Lite, Real-Time, Edge AI.
Abstract:
Human Activity Recognition (HAR) is a technology aimed at identifying basic movements such as walking, running, and staying still, with applications in sports monitoring, healthcare, and supervision of the elderly and children. Traditionally, HAR data processing occurs in cloud servers, which presents drawbacks such as high energy consumption, high costs, and reliance on a stable Internet connection. This study explores the feasibility of implementing human activity recognition directly on embedded devices, focusing on three specific movements: walking, jumping, and staying still. The proposal uses machine learning models implemented with LiteRT (known as TensorFlow Lite), enabling efficient execution on hardware with limited resources. The developed proof of concept demonstrates the potential of embedded systems for real-time activity recognition. This approach highlights the efficiency of edge AI, enabling local inferences without the need for cloud processing.