Vision Sensor for Automatic Recognition of Human Activities via Hybrid Features and Multi-Class Support Vector...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Automated Human Activity Recognition (HAR) is crucial for applications like surveillance and healthcare.
• Existing HAR systems face challenges including poor lighting, varied viewing angles, complex clothing, similar gestures, and limited datasets.

Purpose of the Study:

• To develop a robust HAR system capable of accurately identifying human activities in diverse settings.
• To address the limitations of current HAR systems through effective feature extraction and advanced machine learning models.

Main Methods:

• Extraction of unique key body points and full-body features for distinct activity attributes.
• Utilizing effective spatio-temporal features for human activity identification.
• Employing a Multi-Class Support Vector Machine to enhance classification performance.

Main Results:

• The proposed system demonstrates superior performance in classification, accuracy, and generalization on benchmark datasets.
• Achieved recognition rates of 88.61% on BIT-Interaction, 87.33% on UT-Interaction, 86.5% on NTU RGB + D 120, and 81.25% on PKUMMD.

Conclusions:

• The developed HAR system effectively identifies human activities by leveraging spatio-temporal features and a Multi-Class Support Vector Machine.
• The model's high performance across multiple datasets validates its efficiency and robustness in real-world scenarios.