Non-Invasive Fall Detection Using Computer Vision

Falling is one of the most common, yet devastating problems amongst elderly individuals. Some fall victims can go hours without help if they are injured and cannot reach a phone to call to for help. Our project seeks to develop a prototype that could detect and alert others of a fall. We began by taking 1023 photos in sitting, standing, and laying down(fall) positions. We then fed these pictures to the OpenPose algorithm which outputted the X,Y coordinates of 25 body parts. Those coordinates were then used to train a classifier for falls and non-falls. We tested four different models, a Random Forest, SVM, Neural Network, and Logistic Regression Models all of which performed well. However, the Neural Network had a slightly higher percision and recall. After we had trained a fall detection classifier, we setup a Raspberry Pi 4b with a webcam to make an IoT device. Due computational limits, footage taken from the Pi was sent to a separate computer with a GPU to process the OpenPose algorithm. The outputed body part coordinates were then fed to our pre-trained Neural Network fall classifier. If the fall classifier detects a fall, a python script is activated which alerts others via SMS and email. In total, this prototype cost approximately $85 to build. We believe our method has the potential to be improved upon and possibly implemented in nursing homes or senior communities.