Visual Context-Aware Person Fall Detection

• DOI: 10.48550/arxiv.2404.08088
• Publication Date: 2024-04-11
• Authors: Aleksander Nagaj; Zenjie Li; Dim P. Papadopoulos; Kamal Nasrollahi
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

As the global population ages, number of fall-related incidents is on rise. Effective fall detection systems, specifically in healthcare sector, are crucial to mitigate risks associated with such events. This study evaluates role visual context, including background objects, accuracy classifiers. We present a segmentation pipeline semi-automatically separate individuals and objects images. Well-established models like ResNet-18, EfficientNetV2-S, Swin-Small trained evaluated. During training, pixel-based transformations applied segmented then evaluated raw images without segmentation. Our findings highlight significant influence context detection. The application Gaussian blur image notably improves performance generalization capabilities all models. Background as beds, chairs, or wheelchairs can challenge leading false positive alarms. However, we demonstrate that object-specific contextual during training effectively this challenge. Further analysis using saliency maps supports our observation classification tasks. create both dataset processing API pipeline, available at https://github.com/A-NGJ/image-segmentation-cli.