Research Of Head Pose Estimation Based On Human-Computer Interaction System

With the increasing demand for new human-computer interaction methods, computer vision based human-computer interaction technology has become a hotspot in current research. The head orientations can indicate naturally directions, and head pose estimation makes human-computer interaction more natural and friendly for the elderly and the disabled. Therefore, it is of great practical significance and application prospects to launch research on human-computer interaction based on head pose estimation.A head pose estimation based human-computer interaction approach is proposed in the thesis. Head pose recognition is realized by designing face detection module, face feature extraction module and head pose estimation module. Based on the intelligent wheelchair platform, a head pose estimation based human-computer interaction system is designed and realized.Firstly, a new method combining skin-color segmentation and AdaBoost algorithm is used to detect face in real time and segment multi-poses facial region in the image. To improve head pose recognition rate under changeable illumination, expression, noise and so on, a feature extraction method is proposed. In the proposed method, features can be extracted by taking advantages of both the second order Histogram of Oriented Gradient(HOG) and the Center-Symmetric Local Binary Pattern(CS-LBP). Texture feature is extracted by CS-LBP, and contour feature is extracted by the second order HOG. Both the extracted texture feature and the contour feature are fused to represent face posture more effectively. To reduce the computation time, Kernel Principal Component Analysis(KPCA) is also applied to reduce the dimension of the fused features.Secondly, an improved honey-bee mating optimization random forest algorithm is proposed to construct the facial poses multi-classifier. The proposed algorithm is helpful to enhance the algorithm stability and avoid the local optimum trap resulting from the two introduced random processes. Improved honey-bee mating optimization algorithm is introduced to dynamically change decision tree of random forests, so the diversity of random forest is enhanced, and the accuracy of classification is further improved. As a result, the accurate head pose estimation is achieved. The recognition rates of the proposed head pose estimation method in the FERET and CAS-PEAL-R1 face database reach up to 96.16% and 96.24% respectively, and the method has a good stability.Finally, a head pose estimation based human-intelligent wheelchair interaction system is designed and realized. In this system, the recognition results are converted into control instructions to control the movements of intelligent wheelchair. A number of repetitive and comparative experiments are carried out under different illumination conditions. Experiment results show that the proposed system has good robustness and real-time performance.