中文摘要：

在鲁棒级联位置回归（RCPR）方法中，逐级迭代决策树的结构单一，而且初始化形状机制效率低和不精准。因而提出了一种改进RCPR的人脸对齐方法，采用随机森林的级联回归，由逐级迭代决策树转成结构稳定的并行决策树；同时在测试阶段，采用直方图统计的人脸形状初始化机制，高效地实现了粗定位。改进的方法在多个数据库中验证了可行性，结果表明提出的算法在光照、表情变化、遮挡等情况下能够保持很好的鲁棒性，并且在精度、失败率上都有较大的提高。

英文摘要：

In the robust cascade position regression （RCPR） method, the structure of the decision tree is single, and the efficiency of the initial shape mechanism is low and not accurate. This paper proposes an improved RCPR face alignment method. By using the cascade regression of random forests, the decision tree is transformed into a stable parallel decision tree. At the same time, face shape initialization mechanism with histogram statistics is adopted in the testing stage, so as to achieve efficient coarse positioning. Improved method in multiple database verifies the feasibility. The results show that proposed algorithm can keep good robustness in the light, expression and occlusion. The accuracy and the failure rate are greatly improved.