Media.Player

Computer Recognition of Human Movements from Any Angle

技術優勢

Simplifies classificationEffective with data from any camera and viewpointImproves accuracy

技術應用

Video surveillance Physical therapy Sports enhancement and technique

詳細技術說明

Unlike conventional methods that study human body pose as a whole, this method charts a body pose as a triplet of points. Since any three non-collinear points in the 3D space define a planar surface, a very difficult and highly non-linear problem can be broken down, based on the discovery of projective invariants, into a collection of relatively simple and linear problems. By reducing the original problem of non-rigid body pose estimation to an estimation of poses charted across planes, this innovative method is more widely applicable and accurate than conventional methods.

*Abstract

While typical methods may need exhaustive information from hundreds of examples and different views, this innovation only requires one viewing angle of the subject and one example of each known action in a database. The method simplifies action sequence information by regarding actions as a sequence of point sets or a set of point trajectories in time, using the distilled information to deliver accurate pose identification without needing specific and limited camera parameters. This algorithm only requires these point sets in order to make identifications, without compromising accuracy, increasing overall classification accuracy to 92% for applications in therapy, sports enhancement, and security. For example, security cameras capturing action sequences can use this method to recognize actions and automatically sound an alarm if detecting a security concern. Also, based on recordings of a patientΓÇÖs movement and the comparison to desired actions, a physical therapist can precisely identify problems and deliver specific feedback for improvement.

*Principal Investigation

Name: Hassan Foroosh, Ph.D.

Department:

Name: Yuping Shen, Ph.D.

Department:

國家/地區

美國