湖南大学学报(自然科学版)
湖南大學學報(自然科學版)
호남대학학보(자연과학판)
Journal of Hunan University (Natural Sciences)
2015年
10期
127-132
,共6页
肖进胜%刘婷婷%张亚琪%彭红%鄢煜尘
肖進勝%劉婷婷%張亞琪%彭紅%鄢煜塵
초진성%류정정%장아기%팽홍%언욱진
视频处理%背景建模%混合高斯模型%历史背景
視頻處理%揹景建模%混閤高斯模型%歷史揹景
시빈처리%배경건모%혼합고사모형%역사배경
video processing%background model buildings%Gaussian distribution mixture model (GMM)%history background
针对背景场景重复显现问题,提出了一种基于历史背景的混合高斯模型(His-tory Background-based GMM,HBGMM)。相较于传统的混合高斯模型,该模型对历史背景模型进行标记,并通过判决匹配次数快速调整历史背景模型的学习率。同时对模型权重低于阈值下限历史模型和非历史模型进行区别处理,用该方法更新模型权重从而降低误检率,使历史模型尽量避免误删除。实验结果表明,本文提出的基于历史背景的混合高斯背景模型能够实现记忆背景的功能,从而更快地适应场景的变化,减少前景误判。
針對揹景場景重複顯現問題,提齣瞭一種基于歷史揹景的混閤高斯模型(His-tory Background-based GMM,HBGMM)。相較于傳統的混閤高斯模型,該模型對歷史揹景模型進行標記,併通過判決匹配次數快速調整歷史揹景模型的學習率。同時對模型權重低于閾值下限歷史模型和非歷史模型進行區彆處理,用該方法更新模型權重從而降低誤檢率,使歷史模型儘量避免誤刪除。實驗結果錶明,本文提齣的基于歷史揹景的混閤高斯揹景模型能夠實現記憶揹景的功能,從而更快地適應場景的變化,減少前景誤判。
침대배경장경중복현현문제,제출료일충기우역사배경적혼합고사모형(His-tory Background-based GMM,HBGMM)。상교우전통적혼합고사모형,해모형대역사배경모형진행표기,병통과판결필배차수쾌속조정역사배경모형적학습솔。동시대모형권중저우역치하한역사모형화비역사모형진행구별처리,용해방법경신모형권중종이강저오검솔,사역사모형진량피면오산제。실험결과표명,본문제출적기우역사배경적혼합고사배경모형능구실현기억배경적공능,종이경쾌지괄응장경적변화,감소전경오판。
Classical Gaussian mixture model (GMM)can describe the multimodal state of the video pixels and GMM has certain robustness in dealing with complex scenes,such as slowly changing lighting.However,it still cau-ses false detection because of the change of pixel values in the same position when the background of the scene is re-exposed after being covered.To solve the repetitive background problem,a Gaussian mixture model based on history background (HBGMM)was proposed in this paper.Compared with traditional Gaussian mixture model,this model can quickly adj ust the learning rate by marking the historical background and counting the matched times.We also processed differently between the historical and non-historical model weights lower than threshold to update the model weights to reduce the false detection rate.Experiment results show that the proposed HBGMM can realize the func-tion of remembering the scenes and adapt to the changes of scenes more quickly,thus decreasing the false detection rate.