Foreground Detection Based On Model-sharing Strategy

Foreground detection is a research focus in computer vision. It has been widely used in Automatic Video Surveillance(AVS), automatic driving, video index and classification and motion behavior analysis. In complex scenes, dynamic background, illumination changes, camera jitter and other factors may have adverse effects in foreground detection. How to extract the foreground accurately is the key to the research. In this paper, we focus on dynamic background. By analyzing the temporal and regional variation in dynamic area, we distinguish the dynamic background from the scene first. Then use the mode-sharing strategy to detect the foreground iteratively. The proposed approach leads to a lower false positive rate and has a better detection results.Firstly, in order to distinguish the dynamic background, we present an approach to calculate the background temporal stability based on the wavelet transform. Temporal stability refers to the background changes over time. Different areas have different changes over time. We use Haar wavelet transform to quantify the variation over time so that we can distinguish the dynamic area from the scene.Secondly, we present an iterative mode-sharing strategy as the process of foreground decision. The current pixel is not only compared with its own model, but also may be compared with other pixel’s model which is also in dynamic background. This strategy can make up for the problem that a single model could not provide enough sample information. This way leads to a lower false positive rate.Finally, we apply the mode-sharing strategy to the foreground detection algorithm which is proposed in this paper. It also reference and improve advanced ideas in other algorithms such as Spatial Consistency through Background Samples Propagation in ViBe and Update of the Learning Rate in PBAS. Experiments show that the proposed algorithm leads to a lower false positive rate and higher precision rate. It has a better performance when compared with traditional approach.