Abstract: The traditional finite division can not reflect the distribution of positive and negative samples. In this paper, a new continuous AdaBoost algorithm based on minimum class variance is developed. The new algorithm measures the similarity of the sample through calculating the class variance of every finite division in the process of finite division and selectes the best finite division corresponding to the minimum sum of class variance. Simulations show that the algorithm is of better detection rate and faster convergence.

Key words: face detection, continuous AdaBoost algorithm, training sample, class variance, space partition

摘要： 传统连续AdaBoost算法因等距划分样本空间而无法体现正负样本各自的分布规律。针对该问题，提出一种基于最小类方差的样本空间划分算法。通过计算各种划分方式的类方差，衡量样本的相似性，选取最小类方差和对应的样本作为最佳划分。仿真结果表明，该算法具有较高的检测率和较快的收敛速度。

关键词: 人脸检测, 连续AdaBoost算法, 练样本, 方差, 间划分

CLC Number: 

• TP391