Abstract: In order to detect moving vehicles correctly under the conditions that scene intensity changes rapidly，this paper proposes an online background reconstruction method based on image blocks and K-means clustering.First，the input image is divided into blocks overlapped with each other.Each block uses adaptive match threshold to form several clusters.Then，gain compensation is applied to images of the clusters which are candidates for background to reduce the intensity difference between adjacent blocks.In this process，prior gains are used to make the background intensity close to the input.Finally，current background image is reconstructed with linear blending method to eliminate the edges between blocks.Experimental results show that the method proposed has a good effect in different conditions.