A Human Skeleton Data Optimization Algorithm For Multi-Kinect

As a motion capture device, Kinect can extract human skeleton data (three dimensional positions of body joints) through depth information, which has been increasingly applied in many fields such as action recognition, medical diagnosis and rehabilitation evaluation. But when collecting human skeleton data with a single Kinect, the detection range is limited and stale data exceptions or data missing might occur due to the influence of human body occlusion and environment. These invalid data will directly affect the accuracy of relevant calculation results and hinder the application of Kinect. To solve the above problems, we propose a data acquisition scheme based on multi-Kinect and an optimization algorithm for human skeleton data. First, we use the method based spatial region constraints and K-means clustering to remove the outliers in the human skeleton data collected by each Kinect. Then we convert the data from different perspectives into the same coordinate system and perform adaptive weighted fusion. We finally obtain a stable and valid data sequence, facilitating its application greatly. We design relevant experiments to compare the optimized data with the original data and the results verify the effectiveness of the algorithm.