Development of a vision based pedestrian fall detection system with back propagation neural network

Statistics have shown that most fall events are associated with identifiable risk factors, such as weakness, unsteady gait, medication use, and the environment. Falls can result in abrasions, broken bones, or even death. A real time fall detection system should be developed, which can trigger an alarm people once a fall event occurs. In this study, the proposed scheme obtains image sequences from an interior camera system. The imaged are first used to build up a model of the background using Gaussian mixture model (GMM) with the extraction of foreground images achieved through subtraction. Morphological operations are then used to repair damage to the image and connected-component labeling is used for elimination of noise. From foreground objects, the aspect ratio of the bounding box, the orientation of the ellipse, and the vertical velocity of the center point are extracted for use as input features in a learning algorithm. Fall detection is based on the classification results of learning algorithm using a back propagation neural network.