SemNet: Learning semantic attributes for human activity recognition with deep belief networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) uses sensor data to detect activities.
• Manual creation of semantic features for HAR is expertise-dependent and not scalable.
• Existing methods often rely on handcrafted features, limiting adaptability.

Purpose of the Study:

• To propose SemNet, a deep learning method for automatic semantic feature construction in HAR.
• To address the limitations of manual semantic feature engineering.
• To demonstrate the effectiveness and broader applicability of semantic feature learning.

Main Methods:

• Developed SemNet, a machine learning method based on deep belief networks.
• Employed deep convolutional LSTM for household activity recognition experiments.
• Focused on learning semantic features representative of axial bodily movements.

Main Results:

• SemNet automatically constructs semantic features for HAR.
• SemNet outperforms baseline approaches in activity recognition.
• Learned features correlate highly with manually defined semantic attributes.
• Semantic attribute interpretation is applicable to diverse deep neural network approaches.

Conclusions:

• Deep learning techniques for HAR are semantically meaningful and superior to handcrafted features.
• SemNet offers an automated and scalable approach to semantic feature learning.
• The study provides insights into deep learning methods for Human Activity Recognition.