A New Residual Dense Network for Dance Action Recognition From Heterogeneous View Perception

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Sub-optimal health affects many, increasing interest in physical activities like dance.
• Traditional dance action recognition methods struggle with variations in speed, lighting, occlusion, and background complexity, limiting their robustness.
• Developing accurate and reliable automated dance action recognition is crucial for various applications.

Purpose of the Study:

• To develop an improved residual dense neural network for automatic dance action recognition.
• To enhance the robustness and accuracy of dance action recognition systems.
• To address the limitations of traditional methods in handling variations in action speed, illumination, occlusion, and background.

Main Methods:

• An improved residual dense neural network architecture was employed for feature extraction.
• Convolutional and pooling layers were utilized for initial feature extraction.
• Exponential Linear Unit (ELU) activation function, batch normalization (BN), and Dropout were incorporated to optimize the model, mitigating gradient disappearance, preventing overfitting, accelerating convergence, and improving generalization.
• DenseNet was integrated to enrich and enhance the effectiveness of extracted dance action features.

Main Results:

• The proposed method achieved high recognition rates on three different databases: 99.98%, 97.95%, and 97.96%.
• Experimental results demonstrate the superior performance of the improved residual dense neural network compared to traditional methods.
• The integration of ELU, BN, Dropout, and DenseNet significantly improved model performance and generalization ability.

Conclusions:

• The developed improved residual dense neural network method effectively enhances the performance of dance action recognition.
• The proposed approach offers a robust solution for automated dance action recognition, overcoming the limitations of conventional techniques.
• This study highlights the potential of advanced deep learning architectures for analyzing complex human movements in diverse conditions.