The project aimed at identifying a fall instance in a home-based care setup where in a single image frame there exists more than one visible objects / person for use in homes for the elderly. This would serve as an early notification system to alert health care providers in case a fall incident. This project heavily relies on the publicly available Mask-RCNN implementation by Waleed Abdulla, Matterport Inc. Mask-RCNN. A big thanks to the team for creating this Mask-RCNN library. This project was undertaken as a Master's Thesis Project. The documentation can be found here
To train your own fall detection system, follow these instructions:
- Split your dataset to train and validation.
- Annotate your images with one class which is the fall instance. For this study the annotation tool used was VGG Annotation Tool NB: Please name the class as "fall".
- Place the images in train and validation folders together with the generated JSON files.
- Download the weights [Mask-RCNN coco weights] (https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5) if you will use a pretrained model.
- Train your fall detection model. Highly advisable to use a GPU in training mode.
The images above used for testing were retrieved from here