Human Skeleton Detection and Extraction in Dance Video Based on PSO-Enabled LSTM Neural Network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomechanical Analysis

Background:

• Machine vision is increasingly applied in various fields, including dance education, for human skeleton detection.
• Existing methods suffer from slow recognition speeds and low extraction accuracy, limiting their practical use.
• Accurate human pose estimation is crucial for dance training and analysis.

Purpose of the Study:

• To develop an efficient and accurate method for human skeleton detection and extraction in dance videos.
• To address the limitations of existing technologies in terms of speed and accuracy.
• To leverage particle swarm optimization (PSO) with neural networks for improved pose estimation.

Main Methods:

• A novel neural network architecture integrating particle swarm optimization (PSO-LSTM) was proposed.
• The model was trained and evaluated on benchmark datasets, including MPII and PoseTrack.
• Performance was assessed based on detection speed and extraction accuracy of human skeleton points.

Main Results:

• The PSO-LSTM model demonstrated superior performance compared to existing optimal algorithms on the MPII dataset, with a 3.9% increase in average accuracy.
• On the PoseTrack dataset, the proposed method achieved a 2.3% improvement in average detection and extraction accuracy.
• The results indicate faster detection speeds and higher extraction accuracy for the PSO-based neural network.

Conclusions:

• The proposed neural network based on particle swarm optimization offers significant improvements in human skeleton detection and extraction for dance videos.
• The method exhibits both fast detection speed and high extraction accuracy, making it suitable for educational and training applications.
• This approach advances the application of machine vision in biomechanical analysis and pose estimation within the domain of dance.