Single view head pose estimation system based on SoC FPGA

Head posture estimation is essential for studying human behavior and attention. However, it is still a challenge to achieve good performance in accuracy, speed and robustness for human behavioral detection applications. In this paper, an improved algorithm based on appearance method for head posture estimation is proposed, and a corresponding embedded system based on System-on-Chip (SoC) FPGA is also presented. This implementation utilizes AAM to extract facial feature points and POSIT to estimate the 3DOF head posture. The results show that the system can achieve head pose estimation at a frequency of about 16 fps when the image has a resolution of 640 X480, and the accuracy of the estimation is about 2.7° in mean absolute error.