Adopting Graph Neural Networks to Analyze Human-Object Interactions for Inferring Activities of Daily Living.

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Human Activity Recognition (HAR) is crucial for smart homes and assistive robots.
• Current HAR methods often use pose estimation and object classification from camera frames.
• Leveraging human-object interactions can enhance HAR accuracy and justification.

Purpose of the Study:

• To propose a novel framework for HAR using Graph Neural Networks (GNNs).
• To explicitly analyze human-object interactions for improved daily activity recognition.
• To enhance the inference of environmental objects associated with activities.

Main Methods:

• Developed a framework utilizing Graph Neural Networks (GNNs) to model human-object interactions.
• Encoded correlations from relational data to infer activities and associated objects.
• Evaluated the framework on the Toyota Smart Home dataset.

Main Results:

• The GNN framework achieved 0.88 accuracy in classifying daily activities, outperforming conventional feed-forward networks.
• Incorporating relational data into the GNN improved object-inference accuracy from 0.71 to 0.77.
• Demonstrated superior performance in Human Activity Recognition and object inference.

Conclusions:

• The proposed GNN framework effectively recognizes human activities by analyzing human-object interactions.
• Explicitly modeling interactions significantly improves HAR compared to traditional methods.
• This approach offers a more robust and interpretable solution for activity recognition in smart environments.