A lightweight neural network framework using linear grouped convolution for human activity recognition on mobile devices

Human activity recognition (HAR) has played an indispensable role in ubiquitous computing scenario, from smart homes to game console designing, elderly care, and fitness tracking. It is very hard to manually extract most suitable activity features from raw sensor time series. Due to an obvious advan...

Full description

Saved in:
Bibliographic Details
Published in:The Journal of supercomputing 2022-04, Vol.78 (5), p.6696-6716
Main Authors: Liu, Tianyi, Wang, Shuoyuan, Liu, Yue, Quan, Weiming, Zhang, Lei
Format: Article
Language:English
Subjects:
Algorithms
Artificial neural networks
Compilers
Computer memory
Computer Science
Electronic devices
Feature extraction
Feature recognition
Human activity recognition
Interpreters
Lightweight
Linear transformations
Neural networks
Platforms
Processor Architectures
Programming Languages
Smart buildings
Ubiquitous computing
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Human activity recognition (HAR) has played an indispensable role in ubiquitous computing scenario, from smart homes to game console designing, elderly care, and fitness tracking. It is very hard to manually extract most suitable activity features from raw sensor time series. Due to an obvious advantage, convolutional neural networks (CNNs) that can extract features automatically have been widely utilized for activity recognition. Despite exceptional performance, CNNs are computation-intensive and memory-demanding algorithms because of a large number of internal parameters. As a result, the research attention in HAR implementations over resource-limited embedded platforms has turned to computationally lightweight convolution architectures. In this paper, we offer a contribution in the direction. Simple linear transformations with low cost are combined with a convolution-based HAR classifier to decrease overall computational/memory overhead and, simultaneously, which establishes an efficient classification method with satisfactory performance. Our new method is evaluated against standard convolution-based and residual counterparts, over several popular HAR datasets for algorithm benchmarking. Experimental results verify that, the cheap linear operations can significantly reduce computational and memory cost, and meanwhile producing satisfactory recognition performance, which is able to ensure faster inference on mobile devices. Our new method could be a strong candidate for real HAR implementations on embedded platforms with limited computing resources.
ISSN:0920-8542
1573-0484
DOI:10.1007/s11227-021-04140-5