Improving Person Fall Intelligent Detection Accuracy Using GAN-Augmented Big Data

This paper presents a novel approach for person fall detection that integrates Generative Adversarial Networks (GANs) with a CNN-based fall detection model to enhance accuracy and robustness. By generating synthetic fall scenarios, the GAN addresses the challenge of limited labeled data, introducing varied fall events under diverse conditions such as occlusions, lighting variations, and complex camera angles. These GAN augmented scenarios are combined with real-world fall data to create a more comprehensive training dataset. Additionally, a temporal consistency constraint is applied to improve the model's ability to detect falls as continuous events, thereby reducing both false positives and false negatives. The proposed model was evaluated on a combined real and synthetic dataset, achieving an 86.4% accuracy and an AUC of 89.8%, surpassing both baseline and state-of-the-art models. Results indicate that GAN-based augmentation and temporal consistency constraints significantly enhance the detection of falls, rendering this approach promising for applications in healthcare and safety-critical environments.