Iss2Image: A Novel Signal-Encoding Technique for CNN-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human activity recognition (HAR) traditionally relies on manual feature engineering, demanding expert knowledge and extensive empirical study.
• Deep learning offers automated feature extraction, overcoming limitations of traditional HAR methods.
• Convolutional Neural Networks (CNNs) are well-suited for temporal data like accelerometer (ACC) signals due to their local dependency and scale invariance.

Purpose of the Study:

• To propose an efficient human activity recognition method using deep learning.
• To introduce a novel encoding technique, Iss2Image, for transforming inertial sensor signals into images.
• To develop a CNN model for image-based activity classification.

Main Methods:

• The Iss2Image technique converts time-series inertial sensor data (X, Y, Z axes) into image representations.
• Each axis's data is mapped to a color channel, preserving temporal correlations.
• A CNN model is employed for classifying activities based on these generated images.

Main Results:

• Experimental evaluation on multiple datasets, including a custom smartphone and smartwatch dataset, was conducted.
• The proposed Iss2Image method demonstrated higher accuracy compared to state-of-the-art HAR approaches.
• The image-based classification effectively captures complex patterns in sensor data.

Conclusions:

• The Iss2Image method offers an efficient and accurate approach to human activity recognition.
• Automated feature extraction via deep learning significantly enhances HAR performance.
• Transforming sensor signals into images provides a powerful representation for CNN-based activity classification.