Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (18): 205-217.DOI: 10.3778/j.issn.1002-8331.2102-0043

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Improved Face Tracking Algorithm in Skin-Like Background

XIE Mengda, SUN Peng, ZHANG Zhihao, LANG Yubo, ZHOU Chunbing, SHAN Daguo   

  1. 1.Department of Criminal Science and Technique, Criminal Investigation Police University of China, Shenyang 110854, China
    2.Key Lab of Forensic Science, Ministry of Justice, Shanghai 200063, China
  • Online:2022-09-15 Published:2022-09-15

类肤色背景下的人脸追踪改进算法

解梦达,孙鹏,张志豪,郎宇博,周纯冰,单大国   

  1. 1.中国刑事警察学院 公安信息技术与情报学院,沈阳 110854
    2.司法部司法鉴定重点实验室,上海 200063

Abstract: In the process of face tracking, the CamShift(continuously adaptive mean-shift) algorithm based on the target color features is disturbed by the skin-like background, which leads to the problem of search box offset and abnormal size. An improved algorithm combining the SVM(support vector machines) skin tone segmentation and face tracking monitoring mechanism is proposed. In the Cb and Cr channels of the YCbCr color space, a non-parametric skin color segmentation model and SVM are used to construct a joint skin color segmentation model specific to video standard frames to remove skin colors in video frames in a coarse-to-fine manner. background. The face color histogram in the Cr channel is constructed based on the CamShift algorithm and face tracking is performed. Considering the failure of the joint skin color segmentation model and the color histogram of the CamShift algorithm when the scene or light intensity changes during the tracking process, the pauta criterion is used to detect the outliers of the mean value of the Cr channel in the tracking window. The face detector built by Adaboost(Adaptive Boosting) algorithm is used to initialize the color histogram of CamShift algorithm and reconstruct the joint skin color segmentation model. The experimental results show that the proposed algorithm can track face targets with higher accuracy than the traditional CamShift algorithm in the skin-like background, and it runs faster than the recent tracking algorithms with satisfactory tracking accuracy.

Key words: face tracking, skin color segmentation, support vector machine(SVM), CamShift, outlier detection, AdaBoost

摘要: 针对人脸追踪过程中,基于目标色彩特征的CamShift(continuously adaptive mean-shift)算法受类肤色背景干扰所导致的搜索框偏移及尺寸异常问题,提出了一种结合肤色分割及追踪监测机制的人脸追踪改进算法。在YCbCr色彩空间的Cb、Cr分量内采用非参数肤色分割模型及SVM(support vector machines)构建特定于当前视频序列的联合肤色分割模型,以由粗至细的方式去除视频帧中类肤色背景。随后,在Cr分量内构建CamShift算法色彩直方图并进行人脸追踪。考虑在追踪过程中,当场景或光照强度改变时易出现的联合肤色分割模型及CamShift算法色彩直方图失效问题,采用拉依达准则(pauta criterion)判断追踪窗口内Cr分量均值的异常,当监测到异常值时即判定当前视频帧人脸追踪失败,使用Adaboost(adaptive boosting)算法构建的人脸检测器进行人脸复检并重构CamShift算法色彩直方图及联合肤色分割模型。在OTB-2015目标追踪数据集中进行测试,实验结果表明,所提算法在类肤色背景下相比原始CamShift算法对人脸目标的追踪精度更高;相比近几年的追踪算法则在具有良好追踪精度的同时速度优势明显。

关键词: 人脸追踪, 肤色分割, 支持向量机, CamShift, 异常值检测, AdaBoost