Foreground detection is a significant step of information acquisition in intelligent surveillance. The task is to segment all the moving objects from complex scenes without any false targets and noise interference. This step is a premise of the following steps: object identification

object tracking and behavioral analysis. Due to non-stationary surveillance scenes

foreground extraction becomes a complex task with many challenges. The performance of foreground detection mainly depends on the background modeling algorithm. In order to solve this problem

an adaptive background modeling approach is proposed. This approach is based on a Gaussian mixture model proposed by Stauffer and Grimson. In their approach

each pixel maintains a Gaussian mixture model constituted by Gaussians. Then each Gaussian mixture model is updated by new observe pixel value. However the strategies of updating have some limits

such as fixed Gaussian number

fixed parameters

and fixed learning rate. The proposed approach optimizes updating strategies so as to break these limits. In this approach

each pixel maintains a dynamic Gaussian mixture model

while the number of Gaussians can be controlled dynamically. Further more

an online EM algorithm is applied to the method for estimating the parameters in Gaussian mixture model. At last

several strategies are proposed to control the learning rate of weights. Experimental results show that the foreground object detection approach has good adaptability to complex environments. The foreground object can be detected effectively and rapidly

and the precision and recall ratio of results demonstrate superiority of the method to some related work.