Multiple Kinect Sensor Fusion for Human Skeleton Tracking Using Kalman Filtering

Kinect sensors are able to achieve considerable skeleton tracking performance in a convenient and low-cost manner. However, Kinect sensors often generate poor skeleton poses due to self-occlusion, which is a common problem among most vision-based sensing systems. A simple way to solve this problem is to use multiple Kinect sensors in a workspace and combine the measurements from the different sensors. However, this method creates a new issue known as the data fusion problem. In this research, we developed a human skeleton tracking system using the Kalman filter framework, in which multiple Kinect sensors are used to correct inaccurate tracking data from a single Kinect sensor. Our contribution is to propose a method to determine the reliability of each tracked 3D position of a joint and then combine multiple observations based on measurement confidence. We evaluate the proposed approach by comparison with the ground truth obtained using a commercial marker-based motion-capture system