Leveraging Artificial Occluded Samples for Data Augmentation in Human Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) models struggle with limited and non-diverse training data, leading to overfitting and poor generalization.
• Existing solutions like data augmentation and transfer learning have limitations in addressing real-world challenges.

Purpose of the Study:

• To introduce a novel data augmentation method for HAR that simulates occlusion by removing body parts from skeleton data.
• To enhance the robustness and generalization capabilities of deep learning models for HAR.

Main Methods:

• Developed a new data augmentation technique by artificially occluding skeleton representations, differing from previous rotation-based methods.
• Applied the occlusion simulation to training datasets to increase size and diversity.

Main Results:

• The proposed method effectively increases dataset size and diversity, enabling models to handle a wider range of scenarios.
• Artificially occluded samples enhance model robustness, leading to improved recognition performance, even on non-occluded activities.

Conclusions:

• Simulating occlusion through artificial removal of body parts is an effective data augmentation strategy for HAR.
• This approach improves deep learning model performance and generalization in human activity recognition tasks.