Research On Kinect Skeleton Motion Data Refinement

Human skeleton motion data has been widely used in the fields of game design,film and television production,sports training and medical rehabilitation.People can obtain human skeleton motion data through high-precision motion capture systems,such as Vicon optical motion capture system and Xsens sensor motion capture system.The human skeleton motion data obtained by these high-precision motion data collection devices are abundant in joint information and have high data accuracy.Unfortunately,it is expensive and inconvenient to wear,hence,popularization is difficult.The depth camera Kinect can capture human skeleton motion data in real time.It is cheap and easy to use,but the motion data only contains only 25 skeleton joints and with a lot of noises.If the motion sequence captured by Kinect can be refined to have a similar accuracy level of motion capture data,then the cost of high-precision motion data can be greatly reduced.A stacked-autoencoders-based algorithm is proposed to refine the Kinect skeleton motion data,and a Kinect-oriented high-precision motion data capture system is designed and implemented.It can remove the noises and enhance the details of Kinect motion data simultaneously,and stably optimize the long sequence of Kinect motion data.The main contents include:1)A stacked autoencoder that can process long sequences and a training strategy are proposed.The network consists of stacked bidirectional recurrent autoencoders and perceptual autoencoders.Motion clip pairs of Kinect and motion capture data are used for network training.During the training phase,the mean square error constraints,bone length constraints,smoothness constraints on the output of each bidirectional recurrent autoencoder in the stacked bidirectional recurrent autoencoder,and hidden variable constraints on hidden units of the perceptual autoencoder are utilized.Experimental results demonstrate that the output motion structure is reasonable,smooth and natural,and the proposed stacked autoencoder can effectively remove the noises of Kinect motion data and improve the details of Kinect motion data.2)A high-precision motion data capture system oriented to Kinect v2.0 is implemented based on the proposed stacked autoencoder.The system process long sequences of motion data by means of sliding window.Each motion clipis input into the network in sequence,and the intermediate frame of the network output data is taken as the final refined data.The system can capture Kinect motion data in real-time through the motion recognition tool in Kinect SDK,uses UDP local communication for data transmission between display ends and computing ends,and can convert real-time captured Kinect motion data into high-precision motion capture data for storage,which achieves the purpose of acquiring high-precision motion data based on low-precision and convenient equipment.