Automatic Binary Segmentation of Moving Objects in Video

Contact person: Dmitriy S. Vatolin (

MSU Graphics & Media Lab (Video Group)

Algorithm performs detection and tracking of foreground objects. Each frame of video sequence is subdivided into regions of two types: foreground and background objects. Thus each frame can be represented as combination of two layers. After foreground objects are detected the tracking of them is performed.


The common scheme of the algorithm is shown on the picture below.

Common Scheme of the Algorithm

Algorithm uses complicated foreground-background (FG-BG) discrimination algorithms in order to automatically detect and trace foreground objects. FG objects detection is performed via combined analysis of color and motion information of video stream.

Significant advantage of developed algorithm is the possibility of FG objects detection even in case of very slow motion which is not typical for methods of FG-BG classification. Another advantages of the method are:

Several precision levels of segmentation (from pixel precision to 16x16 pixel block)

  • User assistance is not required during segmentation
  • Adjustable speed/quality trade-off
  • Possibility of tracking several FG objects


Pictures below demonstrate the result of segmentation for frame from test sequence 'bus'. The main obstacle for segmentation here is the abundance of highly textured objects and sophisticated character of motion (bus is moving behind the static fence). But in spite of that developed algorithm produced adequate result:

Original frame Segmentation result

The next example shows the results of segmentation for developed method and University of Florida algorithm (UF). Segmentation results are obtained for test video sequence 'mother & daughter'. This sequence has two obstacles for successful segmentation. The first one is the proximity of colors belonging to different objects. And the second one (obstacle for foreground-background classification) is very slow motion of FG objects. Method of University of Florida produces segments consisting of parts actually belonging to several objects: the blue segment has parts in the area of woman's silhouette, blue segment points are presented around woman's head. However this comparison is not fully correct because algorithms perform segmentation of different types.

Original frame

University of Florida result Proposed method result