基于卡尔曼滤波和随机回归森林的实时头部姿态估计

摘要

Head pose estimation plays an essential role in many high-level face analysis tasks. However, accurate and robust pose estimation is still very challenge with existing approaches. In this paper we propose an accurate head pose estimation method based on random regression forest and Kalman filter with popular RGBD cameras such as Kinect. Firstly, the position of head in the depth image is predicted using Kalman filter, and depth patches are sampled in the prediction area. We then pass these patches through random regression forest to estimate head pose which is considered as the measurement of Kalman filter. Finally, the Kalman filter is used to combine the prediction and the measurement to obtain the final head pose. Compared with the existing random regression for-est algorithm, the experimental results show that this algorithm has faster speed, better robustness and higher ac-curacy.%头部姿态估计在许多高层次的人脸分析任务中起着至关重要的作用,然而准确鲁棒的头部姿态估计仍然是具有挑战性的.针对当前流行的Kinect,提出一种基于卡尔曼滤波和随机回归森林的准确头部姿态估计方法.首先使用卡尔曼滤波在深度图中预测头部的位置,并在预测区域内采样深度块;然后将采样深度块通过已训练的随机回归森林进行头部姿态估计,并将姿态估计值作为卡尔曼滤波的测量值;最后利用卡尔曼滤波结合预测值和测量值得到最终的头部姿态估计参数.实验结果表明,与现有的随机森林算法相比,该方法具有更快的速度、更好的鲁棒性和更高的准确率.