A Comparative Study of Feature Selection Approaches for Human Activity Recognition Using Multimodal Sensory Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Wearable Technology

Background:

• Human Activity Recognition (HAR) is crucial for applications leveraging human movement data.
• Recognizing complex daily activities from unobtrusive wearable motion sensors presents significant challenges due to data diversity and repetitive action sequences.
• Existing methods struggle with the inherent variability in sensor data for activities like cooking or eating.

Purpose of the Study:

• To develop a robust two-level hierarchical method for human activity recognition using wearable sensors.
• To address the challenge of recognizing composite activities by first identifying atomic actions.
• To evaluate feature extraction techniques and classification algorithms for improved HAR accuracy.

Main Methods:

• A two-level hierarchical approach was implemented for human activity recognition.
• Atomic activities were detected from wearable sensor data (smartphone, smartwatch, smart glasses) to obtain recognition scores.
• Composite activities were recognized using these atomic action scores, employing handcrafted features and subspace pooling for feature extraction.

Main Results:

• The proposed method achieved 79% average recognition accuracy using handcrafted features.
• Subspace pooling technique yielded 62.8% average recognition accuracy for composite activities.
• Performance evaluation included various classification algorithms on the CogAge dataset.

Conclusions:

• The two-level hierarchical method effectively recognizes human activities from wearable sensor data.
• The approach demonstrates superior performance compared to existing state-of-the-art techniques.
• This research contributes a novel framework for HAR, enhancing accuracy in complex activity recognition.