Abstract: We propose an improved background extraction algorithm to tackle the problem of low detection accuracy and long ghosting in the classical background extraction algorithm (ViBe) under dynamic background scenes. This method considers the pixel color similarity between spaces and distances in the background model initialization and selects homogeneous pixel dots in the neighborhood to initialize the background model. According to the scene dynamic degree rate adaptive adjustment of each pixel threshold and the background model update, the detection accuracy is improved in dynamic background scenarios. According to the optical flow, movement in the pixel points and the existence of motion target and ghost real prospects can be distinguished, and the background model can be timely modified, thereby eliminating ghosting as soon as possible. Using the change detection test, the improved ViBe algorithm not only can extrat more complete foreground targets, but also can improre the detection accurary compared with original ViBe alogrithm.