Research On Foreground Detection In Videos

Foreground detection is a very important branch in computer vision theory.The accuracy of foreground detection has a large impact on its downstreamprocess, such as object tracking, learning and recognition. The result offoreground detection is influenced by many factors in the videos, some of whichare the shape of foreground object, the environment of the object and theocclusion. Generally speaking, the foreground detection techniques are dividedinto two categories, detection under the stationary camera and the movingcamera.When the video is taken by a stationary camera, the background isstationary and the camera has little or none movement itself. The modeling ofbackground is easy in this situation. The techniques for stationary cameras areoften divided into two: parametric models and non-parameter models, whileparametric models categorize pixels according to static models andnon-parameter models use pixel samples directly for categorization.When the video is taken by a moving camera, it becomes very difficult to model the background. Ego-motion and trajectory categorization combined withtraditional techniques in stationary cameras are often used under this situationOur paper first concludes the overall techniques regarding foregrounddetection in recent years and then study on the following issues:1. Regarding stationary cameras, an improved ViBe approach is putforward, which uses HSV information instead of RGB information ofthe pixels. The model increases the accuracy rate of detection andshows robustness to noise and reciprocation in the videos.2. Regarding moving cameras, a new approach based on optical flows isput forward combined superpixels and MRF-MAP to achieve accurateforeground detection in videos. The method is insensitive to cameramovement and does not require pre knowledge on object trajectories.