人脸位姿协同的多新息抗扰滤波算法

Multi-innovation Disturbance Rejection Filtering Algorithm for Coordination on Face Pose

在偏转角度较大时,人脸特征点的显著性明显减弱,会导致人脸位姿计算结果带有较大噪声.针对这一问题,提出了多新息抗扰滤波算法,将运动人脸与标准人脸模型的位姿变化作为滤波观测量:(1)引入多新息修正滤波估计,利用时间序列的多组观测量估计人脸位姿变化的状态量;(2)实时判断滤波敛散性,根据多新息及时估计观测噪声协方差与过程噪声协方差,调整卡尔曼增益矩阵;(3)建立位姿协同模型,依据滤波后的人脸位姿变化计算相机运动参数,达到相机与人脸位姿协同.在给出试验装置硬件构成的基础上,将本文算法与自适应卡尔曼滤波(AKF)算法进行对比.试验结果表明,在人脸位姿协同系统中,本文算法位姿估计误差小于10 mm,相机协同时间约为25 ms,相较于AKF算法位姿准确度提高23%,协同效率提高30%,能够有效抑制位姿协同中人脸位姿计算所带来的噪声影响,在提高人脸位姿协同系统稳定性的同时,保证响应的实时性.

When the deflection angle is large, great noises are brought to the face pose measurement results because the saliency of facial feature points decreases. To solve the problem, a multi-innovation disturbance rejection filtering algorithm is proposed, in which the pose changes between the moving face and the standard face model are taken as filtering observations.(1) Multi-innovation is introduced to modify filtering estimation, and the states of the face pose changes are estimated by the time series of multi-group observations.(2) The convergence of filtering is judged in real time, and then the noise covariances in the observation and the process are estimated by multi-innovation in time to adjust the Kalman gain matrix.(3) The pose coordination model is established and then the camera motion parameters are calculated according to the changes about face pose after filtering, to realize the coordination between the camera and face pose. On the basis of the hardware structure of the test device, the proposed algorithm is compared with the adaptive Kalman filter(AKF) algorithm. The experimental result shows that the pose estimation error of the proposed algorithm is less than 10 mm and the time of camera coordination is about 25 ms for coordination on face pose. Compared with the AKF algorithm, the pose accuracy is improved by 23% and the coordination efficiency is improved by 30%. The influences of the noises from facial pose calculation in pose coordination can be effectively suppressed by the proposed algorithm. It can improve the stability of the system of coordination on face pose and ensure the real-time response.