Moving Object Detection Algorithm Based On Background Modeling

The moving object detection technology has been a very important research topic in the field of computer vision. In order to extract accurately and completely moving object that the users are interested in from the video sequences, the researchers have done a lot of work. In this paper, we focus on dealing with the effects of the detection results brought by complex scenes in video sequences. Firstly,a method to measure the background complexity is proposed to update the distance threshold and background model rate online, which can achieve robust detection.Secondly, pixel-level background modeling algorithms always have some problems,like incomplete detection results or lack of spatial information. To deal with those problems, a combination with the superpixel segmentation algorithms is proposed.Finally, a method using superpixel features to construct background model is proposed, which can achieve quick and robust detection.Background modeling algorithms have too much noise in detection results, and high error detection rate in dynamic background. To deal with those problems, a moving object detection algorithm based on improved ViBe algorithm is proposed.Firstly, the first frame is used to construct the initial background modeling. Secondly,the sample consensus principle is used to extract the foreground pixels. Moreover, a method to measure the dynamic background complexity is proposed by using the samples' standard deviation. Finally, the value is used to update the background modeling update rate and distance threshold online, which can achieve robust detection. The proposed algorithm is tested on the public dataset CDnet2014,experiments show that this algorithm is robust to dynamic background. The ViBe algorithm is only applicable in the static camera case. To overcome this problem, a method combined with LK optical flow algorithm is proposed, which can reduce the effects brought by background motion. And the improved ViBe algorithm performs well.The pixel-level background modeling algorithms generally have incomplete object, easily missed or false detection results and other issues. To deal with those problems, a combination with the superpixel segmentation algorithms is proposed.The detection image is segmented to several superpixel blocks. Then the superpixel region-filling rate is proposed, which can discriminate whether the superpixel block belongs to the foreground blocks or not. The experiments show this method is robust to background noise, and it can perform well. The superpixel segmentation results have many advantages, like better edge information and controllable superpixel number. Due to those advantages, a method based on superpixel features is proposed.The pixels' mean value in superpixel blocks is uesd as the characteristic value. The superpixel blocks, which are used to construct the background modeling, are located in the same position of initial seed points. The proposed algorithm is tested on the public dataset CDnet2014, experiments show that this method has better performance than pixel-level background modeling algorithms.