In this thesis, new computer vision based techniques are proposed to detect falls of an elderly person living alone. This is an important problem in assisted living.
Different types of information extracted from video recordings are exploited for fall detection using both analytical and machine learning techniques. Initially, a particle filter is used to extract a 2D cue, head velocity,
to determine a likely fall event. The human body region is then extracted with a modern background subtraction algorithm. Ellipse fitting is used to represent this shape and its orientation angle is employed for fall detection. An analytical method is used by setting proper thresholds against which the head velocity and orientation angle are compared for fall discrimination. Movement amplitude is then integrated into the fall detector to reduce false
alarms.
Since 2D features can generate false alarms and are not invariant to different directions, more robust 3D features are next extracted from a 3D person representation formed from video measurements from multiple calibrated cameras. Instead of using thresholds, different data fitting methods
are applied to construct models corresponding to fall activities. These are then used to distinguish falls and non-falls.
In the final works, two practical fall detection schemes which use only one un-calibrated camera are tested in a real home environment. These approaches are based on 2D features which describe human body posture. These extracted features are then applied to construct either a supervised method for posture classification or an unsupervised method for abnormal posture detection. Certain rules which are set according to the characteristics of fall activities are lastly used to build robust fall detection methods. Extensive evaluation studies are included to confirm the efficiency of the
schemes.
History
School
Mechanical, Electrical and Manufacturing Engineering