Leveraging Artificial Occluded Samples for Data Augmentation in Human Activity Recognition

A significant challenge in human activity recognition lies in the limited size and diversity of training datasets, which can lead to overfitting and the poor generalization of deep learning models. Common solutions include data augmentation and transfer learning. This paper introduces a novel data augmentation method that simulates occlusion by artificially removing body parts from skeleton representations in training datasets. This contrasts with previous approaches that focused on augmenting data with rotated skeletons. The proposed method increases dataset size and diversity, enabling models to handle a broader range of scenarios. Occlusion, a common challenge in real-world HAR, occurs when body parts or external objects block visibility, disrupting activity recognition. By leveraging artificially occluded samples, the proposed methodology enhances model robustness, leading to improved recognition performance, even on non-occluded activities.