Background subtractor using mixture of gaussians for moving objects detection. The implementation is based on Stauffer and Grimson algorithm [1].
You only need openCV 3. Building openCV with openMP and the MKL, and compile with -fopenmp (and -O2 and -march too) option is highly recommended to get the best FPS rate.
Constructor parameters are the number of gaussians to model the background and the foreground for one pixel (3 to 5), the learning rate, the minimum weight for a gaussian distribution to be classified as a background and a coefficient to reduce the size of the image for better calculation time. By default, these parameters are set to 3, 1, 0.9 and 0.5. See [1] and [2] to set the learning rate and the minimum weight.
MOGBackgroundSubtraction(int _K = 3, int _downsample = 1, float _a = 0.9, float _T = 0.5);
Just two methods are needed :
void init(std::vector<Mat>& imgs);
Mat createMask(Mat& img);
Give to the first one a vector of cv::Mat to initialize the K gaussians of each pixels (The first N > 10 images in the sequence/video). Then, use createMask to create the mask of the next images. Color black corresponds to the background and white to the foreground.
Images sequences can be found here : https://sites.google.com/site/backgroundsubtraction/test-sequences
main file provide a canvas to use the algorithm. You can use an images sequence, a video or a camera. To do so, you have to provide 2 parameters : "mode" "path"
./mog img directory_with_image/
./mog video video.avi
./mog cam 0
If you have several cameras, change 0 by 1 or 2 or ...
In utils file, you can find a function that reads a directory and put all images names in a vector.
- find a more suitable morphological operation
- automatic set of learning rate and minimum weight
[1] Stauffer C, Grimson W. Adaptive background mixture models for real-time tracking. Proc IEEE Conf on Comp Vision and Patt Recog (CVPR 1999) 1999; 246-252
[2] Thierry Bouwmans, Fida El Baf, Bertrand Vachon. Background Modeling using Mixture of Gaussians for Foreground Detection - A Survey. Recent Patents on Computer Science, Bentham Science Publishers, 2008, 1 (3), pp.219-237.