A Novel Parameter Initialization Technique Using RBM-NN for Human Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Human action recognition is crucial in computer vision but faces accuracy and flexibility challenges.
• Traditional methods struggle with complex actions and large datasets.
• Deep learning offers self-improving models to overcome limitations like overfitting.

Purpose of the Study:

• To propose a novel parameter initialization technique using the Maxout activation function for enhanced human action recognition.
• To improve the accuracy and flexibility of human action recognition models.
• To address limitations in existing state-of-the-art learning models.

Main Methods:

• Human action detection and tracking for spatial-temporal feature extraction from video datasets.
• Training extracted feature descriptors using Restricted Boltzmann Machine-Neural Network (RBM-NN).
• Encoding local features into global features via RBM-NN with integrated forward and backward propagation, followed by Support Vector Machine (SVM) classification.

Main Results:

• The proposed method demonstrated improved recognition rates on benchmark datasets.
• Experimental analysis showed superior performance compared to other state-of-the-art learning models.
• The Maxout activation function and RBM-NN integration effectively enhanced feature learning and classification.

Conclusions:

• The novel parameter initialization technique significantly boosts human action recognition accuracy.
• The integrated approach of RBM-NN and SVM provides a robust solution for complex action recognition tasks.
• This research contributes a more accurate and flexible deep learning model for video-based human action analysis.