SVM directed machine learning classifier for human action recognition network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human action recognition is crucial for surveillance video analysis and public safety.
• Existing methods like 3D CNNs and 2SNNs face computational challenges due to high parameterization.

Purpose of the Study:

• To introduce HARNet, a lightweight residual 3D CNN designed for efficient human action detection.
• To address the computational hurdles of existing action recognition models.

Main Methods:

• Developed HARNet, a lightweight residual 3D CNN based on directed acyclic graphs.
• Created an innovative pipeline for generating spatial motion data from raw video inputs.
• Integrated HARNet with Support Vector Machine (SVM) classifiers using deep-learned features for action recognition.

Main Results:

• Achieved superior performance on benchmark datasets: 2.75% increase on UCF101, 10.94% on HMDB51, and 0.18% on KTH dataset.
• Demonstrated significant performance improvements over state-of-the-art approaches.
• Validated the effectiveness of HARNet's lightweight design and SVM integration.

Conclusions:

• HARNet offers an accurate and computationally efficient solution for human activity recognition in surveillance.
• The combination of HARNet and SVM classifiers enhances discriminative capacity for real-world video analysis.
• This research contributes to advancing surveillance technology for safer and more dependable applications.