Development of a Child Fall Detection System based on YOLOv9 using Transfer Learning from Adult Fall Datasets

Abstract

Falls detection systems are vital for ensuring safety, particularly among vulnerable groups such as children. However, most existing systems are trained using adult-specific datasets, limiting their effectiveness for child fall scenarios due to differences in body size, movement patterns, and fall dynamics. In this study, a child fall detection system was developed using YOLOv9, a state-of-the-art object detection algorithm, with transfer learning applied from adult fall datasets. The system was built through several stages including dataset collection, annotation using Roboflow, model training in Google Colab, and deployment via a user interface created using Gradio. Three experimental setups were conducted: training and testing on adult datasets, testing on child fall images using a model trained solely on adult data, and testing on multi-subject child scenarios. The YOLOv9 model was trained and evaluated at multiple epochs (30, 50, 70, and 100) using standard performance metrics such as precision, recall, F1 score, and mAP. Results showed that while the model achieved high accuracy on adult data (mAP@0.5 up to 0.995), it also demonstrated reasonable performance in detecting child falls, especially at 100 epochs, with a precision of 0.98 and recall of 1.00 in single-child scenarios. Although performance declined slightly in multi-subject scenes, the findings confirm that the developed system can generalize from adult to child fall detection with acceptable accuracy. This study demonstrates the possibility of using transfer learning in child safety as well as providing a cost-effective solution in those settings where a child-specific data is scarce.