Modified Deep Learning Method for Body Postures Recognition

Abstract

The computer vision systems for postures recognitions may possibly be practical solutions for
the health care field as an improvement of healthy aging and as a support to elderly people in
their everyday events. In specific, the fall automatic recognition which attract the attention of
both computer vision and patterns recognitions communities. Most methods based on wearable
sensors, which have the high recognition rates, while some people are unwilling to wear these
sensors. Therefore, alternative methods for example vision based techniques have been
developed. The proposed method is a vision based technique for one-person posture recognition
with the use of convolutional neural network to recognize and classify different classes for the
person posture (e.g., sit, lie, and stand) in each frame (if available), which firstly detects the
person, and for human body shape extraction a background subtraction is applied and each
daily activity is corresponding to extracted shape then estimate the bounded-boxes that
predictable to surround the person body shape. Moreover, the proposed technique results give
the greatest promising solution for indoor monitoring system.