Videosurveillance intelligente pour la detection de chutes chez les personnes agees.

摘要

Developed countries like Canada have to adapt to a growing population of seniors. A majority of seniors reside in private homes and most of them live alone, which can be dangerous in case of a fall, particularly if the person cannot call for help. Video surveillance is a new and promising solution for healthcare systems to ensure the safety of elderly people at home.;We first studied 2D information (images) by analyzing the shape deformation during a fall. Normal activities of an elderly person were used to train a Gaussian Mixture Model (GMM) to detect any abnormal event. Our method was tested with a realistic video data set of simulated falls and normal activities.;However, 3D information like the spatial localization of a person in a room can be very useful for action recognition. Although a multi-camera system is usually preferable to acquire 3D information, we have demonstrated that, with only one calibrated camera, it is possible to localize a person in his/her environment using the person's head. Concretely, the head, modeled by a 3D ellipsoid, was tracked in the video sequence using particle filters. The precision of the 3D head localization was evaluated with a video data set containing the real 3D head localizations obtained with a Motion Capture system. An application example using the 3D head trajectory for fall detection is also proposed.;In conclusion, we have confirmed that a video surveillance system for fall detection with only one camera per room is feasible. To reduce the risk of false alarms, a hybrid method combining 2D and 3D information could be considered.;Concretely, a camera network would be placed in the apartment of the person in order to automatically detect a fall. When a fall is detected, a message would be sent to the emergency center or to the family through a secure Internet connection. For a low cost system, we must limit the number of cameras to only one per room, which leads us to explore monocular methods for fall detection.;Keywords. computer vision, videosurveillance, fall detection, motion detection, tracking, shape analysis, 3D localization.