Head Pose Estimation from Undistorted Wide Field of View (WFoV) Kinect for Socially Acceptable Wheelchair

摘要

Socially acceptable wheelchairs require information about human location and pose to generate a motion that is safe and comfortable to the people around. A WFoV Kinect camera is one of the prominent sensors to obtain such information from a wide coverage of distant targets. However, the task is challenging since the expanded RGB and depth images suffers from distortion, and the head detection were far less than perfect. In this paper, we propose an empirical framework that can alleviate the mentioned problem using laser assisted undistor-tion strategy and depth based segmentation procedure. Initially, Kinect RGB and depth images are corrected using an inverse radial distortion model. Next, the depth data is rectified using a neural network filter based on laser-assisted training. Following that, possible human regions are validated by employing geometrical depth features. The verified depth regions are further processed in its corresponding RGB location, to determine plausible head regions using Haar-like features with the Adaboost classifier. Finally, the obtained head regions are fed to the pose estimation stage that constructed using boosted-based particle filter. Experimental results demonstrate the feasibility of the proposed approach.