The implementation is based on the idea from this paper. Additionally, our version allows quasi-random numbers to be used instead of pseudo-random ones.
You need the following libraries:
- numpy
- scipy
- opencv
The easiest way to install OpenCV is through anaconda. However this method only works if you want to track an object through the webcam. To get opencv through anaconda use:
conda install -c https://conda.binstar.org/menpo opencv3
conda install -c asmeurer pango
If you want to track an object from a pre-recorded video you will have to compile opencv yourself. Instructions to do so can be found here.
If a problem occurs when running the particle filter with quasi-random numbers try running setup.sh to recompile the quasi-random number generator libraries (only works on linux).
Performance is dependent on the video resolution, number of particles and size of particles. The video used for the benchmark had a resolution of 1280x720 pixels, 100 particles were used, with a size of approximately 40 000 pixels. The machine was running Fedora 24, and had an i7-4720HQ. In this setup around 15 frames per second could be processed.
By default the particle filter captures video from the webcam. If you want to run it on a pre-recorded video you have to change the value of two variables:
CAPTURE_FROM = VIDEO
VIDEO_FILE_NAME = "the_absolute_file_name_of_the_video_file"
Just run python particle_filter.py