Design And Implementation Of Head Pose Estimation System Based On Monocular Vision

As the key technique for studying human behavior and attention,head pose estimation has been widely used in face recognition,human-computer interaction,advanced driver assistance systems,intelligent surveillance systems and so on.The traditional intrusive head pose estimation system has the advantage of high precision,but it is cumbersome and expensive.The non-invasive system based on multi-vision is simple in principle,but the field of view is small,the stereo matching is difficult,and the relative position and lighting conditions of each cameras need to be strictly constrained,which seriously restricts its application in real life.Therefore,designing and implementing a head pose estimation system based on monocular vision has certain academic significance and application value.The SoC architecture of ARM+FPGA is chosen as hardware platform,and a non-invasive head pose system in natural light condition is designed based on monocular vision.The main works are as follows: Firstly,SoC technology is used to build the head pose system,and image acquisition is realized on the platform;secondly,the principle of AAM(Active Appearance Model)algorithm for locating face feature points and EPnP(Efficient Pespective-n-Point)algorithm for calculating pose are analyzed and studied,aiming at the shortcomings of the algorithm,optimization and improvement are made to improve the speed and accuracy of head pose estimation;finally,head pose estimation application program is developed by Qt GUI application framework and OpenCV computer vision library,and transplanted to the platform to complete the overall design of the system.The main contributions are as follows:(1)An algorithm scheme for analyzing the posture of human head in 3D space based on a single digital image is proposed,which avoids the problems of small field of view,large computation and harsh conditions for the multi ocular vision system.(2)Aiming at the time consuming problem of locating facial feature points based on AAM algorithm,in the pre-processing process,an efficient Camshift tracking algorithm is introduced to locate the face,and the efficiency of updating the initial position of the AAM model is improved;in the process of AAM fitting,the multi-resolution framework and the fitting approach of "coarse-to-fine" was adopted to reduce the amount of texture information in each iteration,and so as to improve the efficiency of the algorithm.(3)Aiming at the problem that "outbound point" affects the precision of EPnP algorithm,an improved method is proposed to improve the accuracy by reducing the backprojection error.First calculating the re-projection error of the 2D-3D points,and then performs the head pose estimation again according to the first five sets of 2D-3D points with the smallest projection error,which improves the accuracy of the algorithm while ensuring the calculation speed of the algorithm.(4)SoC technology is used to develop the prototype of the head attitude estimation system.The platform has strong expansibility and rich resources,which is beneficial to the expansion and rapid product development of the actual application scene in the later period.The test results show that the system can process 8~9 frames per second,and the average absolute error of yaw,pitch,and roll are 2.64o,3.01oand 1.81orespectively,which is convenient,high-accuracy and stronger adaptability,realizes head poses estimation based on monocular vision and achieves the desired design goal.