摘要
动态人脸跟踪过程中,现有的跟踪算法存在快速运动、遮挡和频繁进出摄像机视野下无法及时判定跟踪漂移导致跟踪失败,而目标再出现时作为新的目标进行跟踪.针对以上难题,提出一种融合跟踪校验和深度学习识别辅助的动态人脸跟踪算法(Kernelized correlation filter with verification and recognition,KCFVR).跟踪算法核心是结合核相关滤波框架,通过跟踪校验算法判定人脸目标是否跟踪漂移导致跟踪失败;在目标重新出现时,结合深度学习网络识别辅助方法判定是否为新目标.实验结果表明:跟踪校验算法及时减少跟踪误差积累,识别辅助算法在跟踪成功率及识别精度上,都取得较优的实验结果,实现同一人脸目标的实时、持续跟踪.
During dynamic face tracking,exiting tracking algorithm cannot track drift in time and cause tracking failure when face region is occluded heavily,fast motion,out of view frequently.When the goal reappears,it is tracked as a new goal.Aiming at the above problems,an improved kernelized correlation filter tracking algorithm is proposed with verification and deep-learning-recognition-aided(KCFVR).Verification algorithm is based on the framework of kernelized correlation filter(KCF),which determines whether the face target is missing.When the target reappears,the proposed method combined with the deep learning network identification assist method can determine whether it is a new target.The experimental results show that tracking algorithm can eliminate tracking error accumulation and it has good result in the tracking success rate and the recognition accuracy.It can implement real-time,continuous tracking of the same face target.
作者
潘今一
王钦
王伟
王亚蒙
刘建烽
PAN Jinyi;WANG Qin;WAGN Wei;WANG Yameng;LIU Jianfeng(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023,China)
出处
《浙江工业大学学报》
CAS
北大核心
2018年第5期481-486,共6页
Journal of Zhejiang University of Technology
基金
国家重点实验室开放基金资助项目(PIL1505)
关键词
肤色检测
人脸检测
核相关滤波跟踪
跟踪校验
识别辅助
skin color detection
face detection
kernelized correlation filter tracking
tracking verification
recognition-aided
作者简介
潘今一(1959—),男,浙江湖州人,教授,研究方向为模式识别与数字图像处理,E-mail:jinyi.pan@eutrovision.com.
