Visualizing Inertial Data For Wearable Sensor Based Daily Life Activity Recognition Using Convolutional Neural...

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Human Activity Recognition (HAR) is vital for healthcare and wellness, particularly in elder care systems.
• Existing HAR methods using handcrafted features struggle with real-world activity complexity.
• Advancements in machine learning have improved HAR accuracy but limitations remain.

Purpose of the Study:

• To develop an efficient wearable sensor-based HAR method.
• To overcome limitations of handcrafted features in current HAR approaches.
• To enhance HAR accuracy for practical applications.

Main Methods:

• Encoding tri-axial inertial data from wearable sensors into color image data.
• Utilizing Convolutional Neural Networks (CNNs) to learn features from the generated images.
• Developing a novel data encoding technique for image-formed representation of sensor data.

Main Results:

• Achieved over 95% recognition accuracy on two challenging activity datasets.
• Demonstrated superior performance compared to other deep learning-based HAR methods.
• Successfully encoded inertial data into discriminative image features for CNN analysis.

Conclusions:

• The proposed image encoding method with CNNs offers an efficient and accurate approach to wearable sensor-based HAR.
• This technique effectively addresses the limitations of handcrafted features in complex activity recognition.
• The method shows significant potential for improving context-aware systems in healthcare and wellness.