| Background modeling is an important step for many video surveillance applications such as object detection and scene understanding. With the progress of computer vision science and technology, background modeling algorithm is so mature that could make perfect detection of foreground object in normal scenes. However, there are some crowded scenes in video surveillance where the proportion of foreground is bigger than that of background in time domain and space domain. In this paper, a novel Pixel-to-Model(P2M) paradigm is presented for background modeling in crowded scenes.In particular, the proposed method models the background with a set of context features for each pixel, which are compressively sensed from local patches. As the previous methods that based on the color of pixel may not be effective in crowded scenes, the P2 M background model is built by the generalized Haar-like feature which comes from compressive sensing theory. This approach using texture and edge information of image patch leads to more accurate background. And it would be more robust for intelligent video surveillance of crowded scenes.Whether a pixel belongs to the background depends on the minimum P2 M distance, which measures the similarity between the pixel and its background model. Moreover, the background updating utilizes minimum and maximum P2 M distances to update the pixel feature descriptors in local and neighboring background models, respectively. That would make the background pixels swallow foreground pixels. The proposed approach is evaluated with foreground detection tasks on real crowded surveillance videos. Experiments results show that the proposed P2 M approach outperforms the state-of-the-art methods both in indoor and outdoor crowded scenes.At last, an expectation calculation approach which based on Bayesian Method is proposed to inpainting our background. The probability of each element comes from minimum P2 M distance. This approach outperforms Gaussian Mixture Model in crowded scenes. |
PDF Full Text Request