摘要
提出一种身份识别模型WMFVR(加权多特征视频识别)。模型分三层:先验层,建立识别群体视频样本库,通过数据发掘可识别(区分)的视频特征点,如身高、肩宽、人脸等,每个特征点聚类建立特征分布,如按身高聚为高、中、低三类;识别层,在相对静止背景下获取连续图像,检测跟踪运动目标,重建人体三维运动,按时间序列识别各个特征点,结合先验知识库,加权决策识别结果;后验层,根据识别结果,更新视频特征分布,刷新先验知识库。与传统视频身份识别相比,WMFVR充分利用连续动态图像下的视频特征,并引入后验机制,提高了视频识别的可应用性。基于WMFVR设计的视频考勤系统识别率为94.872%,表明在固定场景小范围人群的身份识别中具有良好的鲁棒性。
A Weight-Multi-Feature-Video-Recognition (WMFVR) model to recognize stable people with immobile background is presented in the paper. A model consists of three layers .The first layer is a prior repository with recognition people's video features including stature, shoulder width, face etc clustering. The second layer is for video recognition. By capturing continuous images, detecting and tracking moving target, rebuilding three-dimensional movement features, extracting video feature according to time sequence, each feature of target can be recognized orderly .By combing the prior repository and weighting decision, the final recognition decision is the weight of all-feature recognition result. The third layer is posteriori repository for updating video feature distribution, refreshing the prior repository database. Experimental results of WMFVR-based video attendance supervision system show its recognition ratio reaches 95.872%. The WMFVR-based identity recognition system is robust to illumination.
出处
《光电工程》
EI
CAS
CSCD
北大核心
2005年第10期43-46,共4页
Opto-Electronic Engineering
基金
电子信息产业发展基金资助项目([2004]42号)