Digital Signal Processing Reference
In-Depth Information
is also employed to control the contribution of earlier observations and limit the
number of learned subspaces. The proposed algorithm outperforms especially under
sudden and drastic changes in illumination.
To overcome the small foreground object size problem of conventional eigen
background model, Quivy and Kumazawa proposed an improved eigen background
algorithm for large and fast moving foreground objects [
43
]. This proposed method
uses Nelder-Mead Simplex algorithm and a dynamic masking procedure to generate
the background images.
When the scene becomes crowded, it is very difficult for the eigen background
method to model background as some foregrounds may be absorbed into the back-
grounds, leading to severe miss detections and false alarms. Zhipeng et al. proposed a
Selective Eigen background Method for background subtraction in crowded Scenes
[
44
]. The proposed algorithm uses a block-level eigen background algorithm where
the original video frame is divided into blocks and each block is processed indepen-
dently. With the help of blocking strategy, the foreground proportion in the train-
ing samples and the spatio-temporal complexity of the algorithm are significantly
reduced. In order to reduce the absorbtion of foreground object into background,
the algorithm selects the best eigen background for each block to reconstruct its
background, rather than using all the eigen backgrounds in the traditional eigen
background method. To improve this method further, pixel-level selective eigen back-
ground algorithm based on virtual frame is proposed. Virtual frames which contain
no foreground objects are constructed by selecting clean pixels from the video. In the
detection stage each pixel can get the best background reconstruction by selecting
the best eigen background.
In some applications such as sensing web, the privacy information has to be
erased from the image. The existing human detection technique does not works
perfectly and some people cannot be masked correctly so that their privacy cannot
be protected. When a person is detected in a wrong position or not detected, the
mask is overlaid on a wrong position or is not overlaid so that the person is left
unmasked and clearly appeared in the output image. Kawanishi et al. proposed a
background image generation approach to overcome the privacy invasion problem
[
45
]. The proposed algorithm reconstructs the image captured by the camera by
generating a background image without any people and overlaying symbols at the
positions of the corresponding people on the generated image. In this method, even
if the human detection algorithm does not work well, it just causes the rendering of
the symbol on the different position or the lack of the character, but never causes the
privacy invasion.
Background Modelling by Sparse and Redundant Representation
Candes et al. proposed a new method for background modeling [
46
]. He stacked
the video frames as the column of a matrix, then the low-rank component naturally
corresponds to the stationary background and the sparse component captures the
moving objects in the foreground. It allows for the detection of objects in a cluttered
background and offers a way of removing shadows and specularities in images. The
