Action recognition robust to position changes using skeleton information and SVM

Aiming at the problem that the action recognition algorithms based on vision have a high requirement of the background and the human position relative to the sensor, an algorithm which is robust to the position changing of the human is proposed. The Microsoft v2 is used to collect skeleton data and standardize it, then the feature vectors are extracted from the data, at last after the correction of the feature information, the distribution related with the skeleton data is collected and input to the SVM to train. This method can keep stable action recognition when the angle between human torso and Kinect is changed. Experiment result turns out that this position-robust recognition algorithm keeps a recognition accuracy rate above 80%, which is over the state-of-art solutions.