Preprocessing compressed 3D kinect skeletal joints in enhancing human motion classification
Pui Yi Lee, Wei Ping Loh, Jeng Feng Chin · 2016
Human motion classification has been commonly analyzed from 2D and 3D temporal body postures. Previous analyses focused on exergaming, sport science, surveillance and rehabilitation for the betterment of living. A number of recent works had also applied Microsoft Kinect captured motions for its portability and capability convenience to detect RGB images, depth images and 3D skeleton joint coordinates. Nevertheless, Kinect captured data contains inaccuracies which necessitates preprocessing efforts to improve the motion recognition accuracy. To the best of our knowledge, works had rarely detailed the data preprocessing techniques in order to improve the qualities of motion recognition. Few had demonstrated the feasibilities of 2D motion data to represent 3D motions. The importance of data preprocessing role in enhancing the classification accuracy, thus remains a major research concern to date. Therefore, this paper presents the comparisons of human motion data preprocessing analyses on 2D and 3D skeletal joints. Uncertainties observed in raw 3D motion data are identified and filtered followed by the data compression tasks into 2D data. The preprocessing performances at each level are judged using four major classifiers: Bayes, Function, Lazy and Tree aided by the WEKA tool. The approaches are employed on the skeletal joints coordinates of data retrieved from UTKinect-Action Dataset. Our findings demonstrated that preprocessing efforts on the compressed 3D into 2D skeletal joints is comparable to the actual human motion by classification accuracy and at the same time reduces the execution time.