基于均值移动确定性漂移的改进CONDENSATION人脸跟踪

Improved CONDENSATION Face Tracking Algorithm Based on Mean-shift Drift

针对视频序列目标跟踪粒子滤波经典CONDENSATION算法用先验转移概率,即采用一阶或二阶AR模型难以有效进行粒子传播的问题,提出了一种改进的CONDENSATION人脸跟踪算法。首先利用高效的均值移动跟踪器以低廉的计算成本初步进行人脸目标跟踪定位,并用此初步跟踪结果来确定CONDENSATION粒子动态传播模型中的确定性漂移部分,然后只需加入一个较小的随机扩散噪声来完成粒子的传播。由于这样所得的粒子点能较为集中地分布在状态的真实区域附近,因而大大提高了粒子的利用效率。人脸跟踪实验表明,该改进算法的性能明显优于原CONDENSATION方法。

In the classical CONDENSATION for object tracking, a prior transition probability, i.e., first or second order AR dynamic model is used to propagate the particles. However, it results in poor performance frequently. In order to propagate the particles efficiently, an improved CONDENSATION face tracking algorithm based on mean-shift drift is proposed. The approach uses the efficient mean shift tracker to attain coarse location of face target, then uses these results to determine the deterministic drift, finally propagates the particles with a small stochastic diffusion added. Because sampling via the proposed method can always make particles cluster around the true state region, the particles efficiency can be improved greatly. The experimental results of face tracking demonstrate that the performance of proposed algorithm is superior to the standard CONDENSATION.