Depth Data Filtering for Real-time Head Pose Estimation with Kinect

In order to analyze the head motion of pilots in real time and improve tracking performance, we propose a method based on the random regression forest framework to address head pose estimation from depth data captured by Kinect sensors. We present the novel Trinary Annulus Filter and implement Bilateral Filtering using CUDA to process depth data of Kinect, with the purpose of image quality improvement and minimized performance impact. We have evaluated our system on a public database, and it is proved to be more effective after depth data processing and capable of handling large and rapid head rotations, temporary and partial occlusions in performance evaluation. After head pose data are filtered by presented multiple pose estimation method, they are successfully used in flight simulation to drive the rotation of viewpoint.