一种基于粒子滤波的任意姿态头部椭圆轮廓跟踪方法

A method for tracking elliptical contour of arbitrary-pose head based on particle filter

结合粒子滤波思想,给出了一种适用于复杂背景和较远距离下跟踪任意姿态人体头部椭圆轮廓的方法。该方法采用双重随机采样策略,即在预测值附近通过均匀采样产生初始粒子以保证其多样性,在权值更新后对具有较高权值的粒子进行高斯采样以保证重采样过程的快速收敛性。在权值更新过程中,利用分块颜色直方图实现模板椭圆同粒子椭圆之间的颜色匹配,利用最大梯度距离测度(DMG)对粒子椭圆的边缘拟合程度进行度量,最后利用D-S证据思想对上述两种测度进行融合。实验结果不仅验证了此方法对于复杂背景和人体头部的任意姿态具有较强的鲁棒性和有效性,而且证实,此方法也适用于相对较远的距离范围。

Based on particle filters, a head elliptical contour tracking method applicable to complex background, large distance range and arbitrary pose is proposed in this paper. The method applies the strategy of dual stochastic sampling： the uniform sampling around predicted value is utilized to produce initial particles, which can ensure particles＇ diversity; after weights updated, the Gaussian sampling is adopted to resample the particles to achieve high convergence. In weights updating, color matching between template elliptical sub-image and particle elliptical sub-image is implemented by a kind of block histogram, and the distance to maximum gradient point （DMG） on normal line is applied to measure out the particle ellipses＇ similarity to head edges, and finally, by the D-S theory the above two measurements are fused to update the particles＇ weights. The experiments results confirm the method＇s validity, robustness to complex background and arbitrary head pose, and applicability to large distance range.