摘要
基于像素的背景建模方法速度较快但不能很好地描述背景运动,光流能准确描述物体运动但计算量大,难以满足实时的要求.提出一种结合基于像素的背景建模方法速度快以及光流描述物体运动准确优点的背景建模和目标检测方法.具体来说,为静止背景建立传统基于像素的灰度背景模型,为运动背景建立光流背景模型,通过2种背景模型的有效结合快速准确地实现目标检测.实验结果表明,提出的方法建模速度与基于像素背景建模方法相当,同时,又有光流准确描述背景运动的优点,综合性能超越上述2种方法.
The traditional background models based on pixels can not interpret the background motion efficiently although fast in computation. Optical flow can represent object motion accurately, but can not meet the requirements of real time application for computational complexity. In this literature, the traditional background models based on pixels and optical flow are fused with the purpose of combining their advantages, which are used to formulate a novel two model background modeling approach for detecting moving objects fast in computation and accurate in detection. The traditional background models based on pixels are used to model static backgrounds using statistics of pixel intensity, while statistics on intensity, spatial and temporal information of pixels are extracted to generate the optical flow field, which is utilized to model moving ones. Then we can use the two models for moving objects detection fast and accurately. The advantage is that the intensity background model can discriminate foreground from static background fast and accurately, so global optical flow field is not necessary and computational complexity is reduced~ the optical flow background model for moving backgrounds can represent background motion very well, mitigate noise caused by background motion remarkably and detect moving objects accurately and then is superior to the previous two methods. This two model-based background modeling strategy can reduce the noise generated by background motion significantly and detect moving objects fast and robustly, as illustrated in our experiments.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2011年第11期1983-1990,共8页
Journal of Computer Research and Development
基金
国家自然科学基金项目(90924026)
国家"八六三"高技术研究发展计划基金项目(2007AA01Z338
2008AA01Z121)
关键词
背景建模
光流法
目标检测
混合高斯
双模型
background modeling
optical flow
object detection
mixture of Gaussian
two-model