Enhanced Video-Based Fall Detection Using 3D Human Pose Estimation and Advanced Generative Adversarial Networks: Integrating WGAN Data Augmentation with a Hybrid 3D CNN–GRU Architecture

• Qasim Mahdi Haref,
• Jun Long,
• Zhan Yang,

Vol. 19, No. 12, December 31, 2025
10.3837/tiis.2025.12.009, Download Paper (Free):

• Wasserstein GAN
• Fall Detection
• CNN-GRU
• video surveillance
• Human Activity Recognition
• 3D pose estimation
• healthcare AI

Abstract

The leading cause of injury among the elderly is falls, although video-based detection is constrained by the lack of labeled data, environmental diversity, and most often false alarms. This paper proposes a deep learning model that combines realistic data augmentation with spatiotemporal motion modeling to overcome them. A Wasserstein generative adversarial network (WGAN) is used to build 3D sequences of human skeletons to reduce the issue of class imbalance and enhance generalization of the model. The analyzed augmented data is then analyzed with a hybrid 3D convolutional-recurrent model which is simultaneously trained on spatial configurations and temporal motion cues. It includes a single-shot detector (YOLOv8) that is used to localize persons; it guarantees the accuracy of region proposals and minimizes the background noise prior to the pose estimation. The suggested method is tested on Le2i fall detection benchmark, with a test accuracy of 97.33, which is much higher than the baseline that is trained on real data only (89.70). The findings of experiments prove the versatility of the model and its capability to separate falls and everyday actions in the complicated real-life situations. A scalable and reliable fall detection based on the generative data expansion and efficient use of spatiotemporal learning designed in the framework makes it feasible to apply to healthcare and surveillance. In general, the study has shown that incorporation of synthetic motion generation with sophisticated deep learning models can significantly improve the performance, generalization, and functionality of vision-based fall detection systems.

Statistics

Cite this article