A dynamic light image enhancement algorithm using generative adversarial network for group activity recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human Activity Recognition (HAR) is crucial for tracking daily activities, often extended to Group Activity Recognition (GAR) for complex scenarios.
• Existing GAR models face limitations including variable image quality, dynamic environmental conditions (e.g., lighting), and computational demands from large datasets and complex architectures.

Purpose of the Study:

• To enhance the performance and robustness of Group Activity Recognition (GAR) models.
• To address limitations in image quality, dynamic environments, and computational complexity in GAR.

Main Methods:

• Proposed a dynamic light image enhancement generative adversarial network (GAN) for improved image quality.
• Introduced a multi-input image-enhanced generative adversarial network (MIIEGAN) for generating synthetic training data.
• Developed a guided asymmetric depthwise separable convolution (GA-DSC) to optimize model complexity and performance.
• Evaluated algorithms on benchmark datasets (The Volleyball Dataset, The Collective Dataset) using multiple deep learning backbones.

Main Results:

• The proposed methods significantly enhance image quality and GAR model performance.
• Achieved superior results compared to existing methods, GAN variants, and convolutional techniques.
• Demonstrated robustness in dynamic light conditions, crucial for real-world indoor and outdoor applications.

Conclusions:

• The novel approach effectively tackles key challenges in GAR, leading to significant performance improvements.
• The developed techniques offer a more robust and efficient solution for real-world group activity recognition applications.