Background Subtraction is a key element in many algorithms used in
the context of image sequence analysis. The distinction between image
background and foreground is, for instance, a meaningful cue for the
detection and tracking of pedestrians in video surveillance.
The goal of this exercise is to obtain an advanced understanding of the
approaches for background subtraction presented in the lectures. The
exercise involves the modelling of the background and the decision for
every pixel whether it belongs to the foreground or to the background.
The background model to be used for this exercise is the "Single
Image sequence of a surveillance camera in a station
Implementation of the sequential estimation for the parameters
of a single Gaussian (mean and variance).
Implementation of the background/foreground labelling process
for every pixel using the variance of the background model
(threshold at 2.5σ).
Simple implementation of a noise reduction in the labels using
Evaluate the given image sequence with two different values for
the learning rate 1 2 ,where 1/50, 1/1400
Questions to be answered in written report:
Apply the implementation of the background subtraction
algorithm using a single Gaussian as a background model to the
given image sequence. Display and watch the results to get an
impression of the effects that can be observed. Describe the
results you got after applying your implementation to the given
image sequence concisely. Observations about your results,
e.g., how results change with different learning rates, in which
part of the image wrong labelling occurs more often, are also
required to be summarized.
Chose two meaningful image frames for which you show
a) the background mean image,
b) the background variance image,
c) the difference image of mean background and current frame,
d) the binary foreground/background mask
for a learning rate 1/50. Indicate the frame numbers in the caption of the figure.
Name the advantages and drawbacks of the single Gaussian model.
Discuss the result of your implemented method when facing difficult cases (consider shadows, reflections, people/things standing still, etc.) and interpret the reason for 1 or 2 cases where pixels are wrongly labelled.