Combining Supervised and Unsupervised Learning Algorithms for Human Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human activity recognition (HAR) is a key area in AI, with recent advances using deep learning to model spatio-temporal dependencies.
• Existing methods often rely solely on supervised or unsupervised learning, presenting limitations in real-time performance and adaptability.

Purpose of the Study:

• To propose a novel, hybrid approach for human activity recognition using skeleton data.
• To achieve high accuracy and real-time performance by combining unsupervised and supervised learning techniques.

Main Methods:

• A two-stage framework integrating unsupervised clustering for activity grouping and graph convolutional networks (GCNs) for classification.
• Exploration of various clustering techniques and data augmentation strategies to optimize the training process.

Main Results:

• Achieved 90.22% Top-1 accuracy on the NTU-RGB+D dataset, an improvement of approximately 9% over baseline GCN methods.
• Maintained comparable inference times and model parameter counts to existing state-of-the-art approaches.
• Demonstrated robust and fast extension of activity recognition capabilities to new classes without full model retraining.

Conclusions:

• The proposed hybrid approach offers a significant advancement in human activity recognition accuracy and efficiency.
• The framework's modularity allows for rapid adaptation and scalability to new activities, making it practical for real-world applications.