Project 1 Motion Detection Using Simple Image Filtering

Computer Vision I
Project 1
Motion Detection Using Simple Image Filtering
The purpose of this project is to get familiar with manipulating images and image filtering techniques.
1. Motion Detection
In this project you will explore a simple technique for motion detection in image sequences captured with a
stationary camera where most of the pixels belong to a stationary background and relatively small moving
objects pass in front of the camera. In this case, the intensity values observed at a pixel over time is a
constant or slowly varying signal, except when a moving object begins to pass through that pixel, in which
case the intensity of the background is replaced by the intensity of the foreground object. Thus, we can
detect a moving object by looking at large gradients in the temporal evolution of the pixel values.
2. Project Requirements:
(a) Write a program to implement the above ideas. Sample videos to test your program are available in
blackboard. The basic outline of your program should be as follows:
i. Read in a sequence of image frames and make them grayscale.
ii. As enough frames are available, apply a 1-D differential operator at each pixel to compute a temporal
derivative.
iii. Threshold the absolute values of the derivatives to create a 0 and 1 mask of the moving objects.
iv. Combine the mask with the original frame to display the results.
(b) You should explore the following variations on the above basic outline:
i. For the temporal derivative filter try a simple 0.5[−1, 0, 1] filter and a 1D derivative of a Gaussian
with a user defined standard deviation tsigma. Compare results running these filters; for the
Gaussian filter try increasing values of tsigma. Discuss your results.
ii. Try applying a 2D spatial smoothing filter to the frames before applying the temporal derivative
filter. For the spatial smoothing filter try 3 × 3, 5 × 5 box filters and 2D Gaussian filters with a
user defined standard deviation ssigma. Run the program using different amounts of smoothing.
Discuss your results. Are there values that make it work better? Explain what you mean by better.
iii. Vary the threshold to get the mask and see what works best. Design a strategy to select a good
threshold for each image. Hint: Most pixels belonging to the background change little and thus
their temporal gradients are close to zero. You can model these values as Gaussian zero mean noise
and estimate the standard deviation of this noise.
(c) Write a report. The report should include:
i. Abstract, description of algorithms, experiments, values of parameters used, observations and conclusions.
ii. Representative images that you tested the algorithms with as well as representative outputs showing
both “good” and “bad” results.
iii. An appendix with your code