Chan,K.L.
Abstract:
Background modeling is an important issue in video surveillance. A sophisticated and adaptive background model can be used to detect moving objects which are segregated from the scene in each image frame of the video via the background subtraction process. Many background subtraction methods are proposed for video acquired by a stationary camera, assuming that the background exhibits stationary properties. However, it becomes harder under various dynamic circumstances – illumination changes, background motions, shadows, camera jitter, etc. We propose a versatile background modeling method for representing complex background scenes. The background model is learned from a short sequence of spatio-temporal video data. Each pixel of the background scene is represented by samples of color and local pattern. The local pattern is characterized by perception-inspired features. In order to cater for changes in the scene, the background model is updated along the video based on the background subtraction result. In each new video frame, moving objects are considered as foregrounds which are detected by background subtraction. A pixel is labeled as background when it matches with some samples in the background model. Otherwise, the pixel is labeled as foreground. We propose a novel perception-based matching scheme to estimate the similarity between the pixel and the background model. We test our method using common datasets and achieve better performance than various background subtraction algorithms in some image sequences.