Skeleton tracking based complex human activity recognition using kinect camera

Abstract

This paper presents a new and efficient algorithm for complexhuman activity recognition using depth videos recorded from a singleMicrosoft Kinect camera. The algorithm has been implemented onvideos recorded from Kinect camera in OpenNI video file format (.oni).OpenNI file format provides a combined video with both RGB and depthinformation. An OpenNI specific dataset of such videos has been createdcontaining 200 videos of 8 different activities being performed by differentindividuals. This dataset should serve as a reference for futureresearch involving OpenNI skeleton tracker. The algorithm is based onskeleton tracking using state of the art OpenNI skeleton tracker. Variousjoints and body parts in human skeleton have been tracked and the selectionof these joints is made based on the nature of the activity beingperformed. The change in position of the selected joints and body partsduring the activity has been used to construct feature vectors for eachactivity. Support vector machine (SVM) multi-class classifier has beenused to classify and recognize the activities being performed. Experimentalresults show the algorithm is able to successfully classify the set ofactivities irrespective of the individual performing the activities and theposition of the individual in front of the camera.