Here lies my university course project
I had an idea to increase accuracy and robustness of a famous KCF tracker. And I thought maybe a Kalman filter can do its filtering thing and place a bounding box more accurately. But I didn't want to select parameters of Kalman filter by hand, so I implemented some sort of genetic algorithm to do it for me.
I used generic Kalman filter algorithm. The state vector was as follows: (x1, y1,), where (x1, y1) - is the center of the bounding box.
Genetic algorithm selected elements for B, u, S, R and A matrices. I used f-measure with beta = 3 as a fitness function. Other parameters of the fitness function were accuracy and robustness. They were calculated plain VOT way, although I used DIOU instead of IOU, and omit less then 10 frames before calculation of the metrics.
Here:
-1.04996 -0.35197 2.14865 1.81949 1.0531 0.379274 -0.972355 -4.83614 -0.664763 1.58524 3.81422 0.125361 -0.805518 1.51277 -2.40536 -3.02137 2.54712 0.250593 0.310221 0.0344269 -0.52078 2.17617 0.0965927 0.910537 0.0220096 0.745542 -1.34083 0.2846 -1.66556 -1.72201 -0.333888 2.63724 1.7406 -2.1478 -0.25823 0.908437 3.43149 -0.00610241 -0.767901 0.0167775 1.48063 0.0126195 0.167397 -0.0533417 -0.0268063 -0.04304 -0.0294877 -2.30189 0.188723 -0.300192 1.17634 1.12888 0.0306505 0.0520604 -0.249356 -3.18886
Compiled and tested with OpenCV 4.2.0
pip install git+https://github.com/votchallenge/vot-toolkit-python
# if you have the same problem with numpy version that I did, then run this command:
# pip install numpy=1.21
mkdir vot2017
vot initialize vot2017 --workspace vot2017
# After loading process finishes,
# you'll have directory with 60 sequencesBuild train code and test code
mkdir build && cd build
cmake ..
make -j4# Start train
KCF# Or run test on a webcam stream.
# Genome already included in code (look at src/runtracker.cpp)
run_tracker
# When you run run_tracker,
# press B and select BBox to track, than press Enter