Background Modeling with Freely Moving Camera

Abstract

Detecting moving objects with a moving camera is a challenge task due to camera motion, parallax effect, registration error and etc. Currently, a lot of techniques detect moving objects by establishing a panoramic background model for the video sequence captured by moving camera. In this process, motion compensation is usually used to registrar the pixel between current frame and panoramic background. However, most of methods don't consider the registration error (caused by inadequate motion model, incorrect parameter estimation and computation error etc.) and usually update background model directly in the location after registration. Since it is difficult to realize the motion compensation with sufficient pixel accuracy, the pixel-wise decision for foreground or background through background model will fail. In this paper, we use the spatial distribution of Gaussians (SDG) model to reduce the registration error caused by motion compensation. According to this model, we only assume the alignment between current frame and background image is approximate, we will find the real correspond location for each current frame pixel in the neighborhood of the predicted location in background image. Based on this model we proposed a new algorithm to model panoramic background for video sequences captured by the freely moving camera. The proposed method and the original method are evaluated and compared based on six video sequences taken by PTZ camera and handheld camera. Compared with traditional method, the proposed method is works well both in detection accuracy and the effect of background modeling.