Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (14): 175-179.DOI: 10.3778/j.issn.1002-8331.1801-0369

Previous Articles     Next Articles

Face tracking using siamese convolutional neural networks

WU Hanzhao   

  1. School of Software, Tsinghua University, Beijing 100084, China
  • Online:2018-07-15 Published:2018-08-06



  1. 清华大学 软件学院,北京 100084

Abstract: Due to illumination variation, occlusion and other reasons, face tracking has challenges in the real world. A method based on Convolutional Neural Networks(CNN) for face tracking, which adopts improved siamese CNN, is presented. Taking paired image areas as input, the distance as output which evaluates the similarity of that image areas pair, the siamese CNN is trained end-to-end on the OTB dataset. Further, the bounding-box regression is applied to improve target localization accuracy. The results show that the proposed siamese CNN is superior to basic convolutional neural networks and achieves better performance on face tracking.

Key words: deep learning, convolutional neural networks, face tracking, bounding box regression

摘要: 由于光照、遮挡、尺度变化等原因,在真实多变场景下完成人脸追踪面临挑战。探究了基于卷积神经网络(CNN)的人脸追踪,将基本的卷积神经网络改进为孪生神经网络,在OTB数据集上采用端到端的方式,以成对图像区域作为输入,输出两者距离,通过距离评估图像区域相似性;加入边框回归算法(bounding box regression)微调追踪结果。实验结果表明,改进后的神经网络优于传统的卷积神经网络,能达到更好的人脸追踪效果。

关键词: 深度学习, 卷积神经网络, 人脸追踪, 边框回归