Video action recognition collaborative learning with dynamics via PSO-ConvNet Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Human Action Recognition (HAR) is challenging due to the need for temporal feature analysis, which standard Convolutional Neural Networks (ConvNets) struggle with.
• Existing methods often fail to capture the dynamic, temporal aspects crucial for accurate video-based action classification.

Purpose of the Study:

• To propose a novel dynamic Particle Swarm Optimization-Convolutional Neural Network (PSO-ConvNet) model for enhanced Human Action Recognition (HAR) in video sequences.
• To integrate ConvNets with advanced temporal models like Transformers and Recurrent Neural Networks to capture spatio-temporal dynamics.
• To evaluate the effectiveness of collaborative learning within the proposed framework.

Main Methods:

• Developed a dynamic PSO-ConvNet model where neural network weights act as particle positions in phase space, facilitating shared learning.
• Integrated ConvNets with Transformer and Recurrent Neural Network architectures to process temporal information in videos.
• Employed collaborative learning strategies, comparing them against individual learning paradigms.

Main Results:

• Achieved significant accuracy improvements of up to 9% on the UCF-101 dataset.
• Demonstrated superior performance of collaborative learning over individual learning on larger datasets like Kinetics-400 and HMDB-51.
• Validated the model's effectiveness in capturing spatio-temporal dynamics for HAR.

Conclusions:

• The dynamic PSO-ConvNet model offers a promising advancement for Human Action Recognition by effectively integrating Convolutional Neural Networks with temporal modeling.
• Collaborative learning within the PSO-ConvNet framework significantly boosts performance in HAR tasks.
• The proposed method provides a robust solution for analyzing complex human actions in video sequences.