A union of deep learning and swarm-based optimization for 3D human action recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human Action Recognition (HAR) is crucial for applications like surveillance and healthcare.
• 3D skeleton data offers a simplified and cost-efficient approach to HAR.
• Existing methods may benefit from enhanced feature extraction and optimization.

Purpose of the Study:

• To propose DSwarm-Net, a deep learning and swarm intelligence framework for HAR using 3D skeleton data.
• To develop an image-encoding method for skeletal features to simplify classification.
• To optimize feature representations using the Ant Lion Optimizer for improved accuracy and reduced dimensionality.

Main Methods:

• Extracted four skeletal features: Distance, Distance Velocity, Angle, and Angle Velocity.
• Encoded features into images and stacked them depth-wise.
• Utilized a modified Inception-ResNet Convolutional Neural Network for classification.
• Applied the Ant Lion Optimizer to refine extracted deep features.

Main Results:

• DSwarm-Net achieved competitive results on UTD-MHAD, HDM05, and NTU RGB+D 60 datasets.
• The image-encoding approach simplified the action classification task.
• Feature optimization effectively reduced dimensionality and removed non-informative features.

Conclusions:

• DSwarm-Net demonstrates superior performance compared to state-of-the-art HAR models.
• The integration of deep learning and swarm intelligence offers a powerful approach for HAR.
• The proposed method provides an effective and efficient solution for action recognition from 3D skeleton data.