A novel method for user-defined human posture recognition using Kinect

Human posture recognition is very critical in human computer interaction studies. With the release of Microsoft Kinect sensor, there has been an increasing interest in using Kinect for vision based human posture recognition as user's skeleton information can be precisely inferred from the depth images generated by Kinect. In this paper we proposed a novel human posture recognition method using Microsoft Kinect sensor, which can automatically identify any user-defined postures. Skeleton information inferred from depth image of user's posture was utilized to generate 9 features representing specific body parts such as forearm, thigh, etc. These features are fed into SVM to generate posture-learning models that are then used to identify pre-defined postures. Totally 22 different postures including body, arm, leg postures were collected and PCA analysis demonstrated they were in general well separated in the feature space. Further performance evaluation using 10-fold cross-validation showed a final overall accuracy of 99.14% was successfully achieved in the test including all postures, indicating the outstanding capability of this proposed methods.