Pose-Based Fall Detection Across Views: Boosting Generalization with Synthetic Data and Data Augmentation

• Eduardo Façanha Dutra Unifor
• Arthur Cavalcante e Silva Unifor
• Maria Andréia Formico Rodrigues Unifor

Resumo

Fall detection models must generalize to diverse, uncontrolled environments for effective real-world deployment, yet achieving this remains a significant challenge in vision-based systems. This paper investigates techniques to enhance generalization, initially focusing on cross-view performance as a critical step toward broader robustness across datasets and settings. We leverage 2D pose data, synthetic training samples, and data augmentation to address this issue. A synthetic dataset was generated using Unity, simulating six 3D fall scenarios captured from ten virtual cameras with precise keypoint annotations. Experiments utilized a public, multi-view real-world dataset, training models on one camera and testing on others to simulate generalization to unseen perspectives. The baseline results showed notable drops in across test views, revealing overfitting to specific viewpoints. We found that lightweight data augmentations—such as flipping, rotation, and Gaussian noise—applied to pose sequences markedly improve cross-view generalization. Combining these with synthetic data further enhances performance, delivering consistent results across all cameras. The most effective strategy, integrating both approaches, achieved gains of up to 54.4% in over the baseline. These results emphasize the value of training diversity, via augmentation and synthetic data, in crafting robust, viewpoint-independent models. By advancing cross-view generalization, this work establishes a foundation for future efforts targeting broader generalization to diverse datasets and real-world environments, providing practical insights for assistive technology deployment.