(a)
(b)
Background estimation is a key step in video surveillance systems. The objects in motion can be easily grabbed by subtracting the foreground image from the background estimated.
The simplest way of background estimation is using a pre-recorded background image. This method works well for well controlled setup of environments, e.g., an assembly line.
For outdoor setup of vehicle surveillance system, using a static background image is apparently not appropriate since the background will change due to different lighting conditions and weather conditions.
A statistical approach was applied so that we can estimate the background dynamically and correctly in spite of the changes of the time and the weather. N background images were kept in a queue of buffers. Once a while, the oldest sample will be replaced by the current background image to ensure that all the background samples are "fresh". For each pixel, the median value of the N samples is considered as the current estimation. Figure 1(a) shows a moving vehicle and Figure 1(b) shows the difference between this image and the estimated background.
|
(a) |
(b) |
Figure 1. (a) A moving vehicle in the Region of Interest marked by a red polygon ; (b) The detected motion object using background subtraction.
As we can see, the shadow cast by the minivan is also considered as object in motion and it is important to remove the shadow from the detected motion region.
The color value of a pixel under shadows changes. The changes can be decomposed into brightness and chromaticity components. The color value changes of pixels under shadows contribute mainly to the brightness components. This observation leads to the approach we used to remove shadows. In Figure 2(a), those pixels with color value changed are marked as red if the changes are mainly changes on the brightness components. Therefore they are shadow suspicious. In Figure 2(b), only those pixels having their chromaticity components changed are marked and they make up a more reliable region of the object in motion.
(a) |
(b) |
Figure 2. (a) Pixels with changes on the brightness components are marked as red; while those with changes on the chromaticity components are marked as blue. Although the color value of those red pixels changed, but they are shadow suspicious; (b) The pixels with changes on the chromaticity components make up a more reliable region of the object in motion.