Neural Networks for Automatic Posture Recognition in Ambient-Assisted Living

Area of Science:

• Artificial Intelligence
• Computer Vision
• Human-Computer Interaction

Background:

• Human Action Recognition (HAR) is crucial for Ambient Assisted Living (AAL).
• Frail individuals require systems promoting autonomous, safe, and secure living.
• AI-powered monitoring systems can detect dangerous situations by classifying human postures.

Purpose of the Study:

• To develop and compare AI models for Human Action Recognition (HAR) in an Ambient Assisted Living (AAL) context.
• To evaluate the performance of Multi-Layer Perceptron (MLP) and Long-Short Term Memory (LSTM) networks for posture classification.
• To optimize feature selection and model architectures for improved HAR accuracy.

Main Methods:

• Utilized skeleton data from four Kinect One systems for multi-angle scene recording.
• Employed Artificial Intelligence (AI) solutions, specifically MLP and LSTM Sequence networks.
• Implemented SVM and genetic algorithms for feature selection, followed by hyperparameter optimization.

Main Results:

• The best performing LSTM model achieved 85.7% accuracy.
• The best performing MLP model achieved 78.4% accuracy.
• LSTM demonstrated better suitability for individual class recognition.

Conclusions:

• LSTM networks offer superior performance for Human Action Recognition in AAL compared to MLP.
• The developed system effectively classifies human postures for detecting dangerous situations.
• AI-driven HAR systems hold significant potential for enhancing safety and autonomy in assisted living environments.