Real Time Human Action Recognition using Kinematic State Model
Geetanjali Kale, Varsha Hemant Patil · 2016
Human action recognition has tremendous applications in interdisciplinary domain and it's challenges kept researchers busy worldwide. Variety of applications gave rise to different representation and recognition methods. Posture can either be represented by shape features or skeleton features. We have represented action using sequence of postures and skeleton features are used for posture representation. Proposed work recognizes Yogasana from real time video. Yogasana is a type of exercise, in which specific sequence of the postures needs to be performed. Regular practice of it shows tremendous benefits in physiological and psychological disorders. System uses skeleton data of twenty human joints for representation of posture. Asanas are represented by kinematic state model using skeleton data provided by Kinect sensor. System is tested on 120 real time video sequences captured on four different subjects performing three asnas. Our system gave 96 % recognition rate.