Appearance based tracking

Lucas-Kanade tracking

The original (non Pyramidal) tracker with fixed window (its size doesnt update) is implemented.

To run Lucas Kanade tracking type in Terminal:
python3 lucas_kanade_tracking.py [--roi roi] [--dpath path_to_images]
The process will start. After the new region of interest was found, it is printed in console.
To stop tracking press Ctrl+C.

Was tested on 3 datasets from http://cvlab.hanyang.ac.kr/tracker_benchmark/datasets.html
The results of tracking can be compared with groundtruth_rect.txt which is provided in archive downloaded by the link.

1. Coke
   Run with: python3 lucas_kanade_tracking.py --roi 298 160 48 80 --dpath 'Coke/img/' The tracker should track the can, but in 6 iterations after the start finds a hand and tracks the hand.

2. DragonBaby
   Run with: python3 lucas_kanade_tracking.py --roi 160 83 56 65 --dpath 'DragonBaby/img/' Completely misses a baby when it starts moving fast.

3. Football
   Run with: python3 lucas_kanade_tracking.py --roi 310 102 39 50 --dpath 'Football/img/'
   On this dataset tracker performs the best because the region of interest changes a little on each iteration.

Meanshift tracking

The original meanshift tracker is implemented and compared with OpenCV Meanshift and Camshift.

To run Meanshift tracking type in Terminal: python3 mean_shift.py [--roi roi] [--dpath path_to_images] The process will start. Three tracking windows are shown on image:
- Red - custom Meanshift
- Blue - OpenCV Meanshift
- Green - OpenCV Camshift

Was tested on 2 datasets from http://cvlab.hanyang.ac.kr/tracker_benchmark/datasets.html
The methods don't work properly with small window, so the window should be larger. To track the process in details change the last line in show_image_with_rect function to: cv2.waitKey(0) The results of tracking can be compared with groundtruth_rect.txt which is provided in archive downloaded by the link.

1. Coke
   Run with: python3 mean_shift.py --roi 270 160 80 80 --dpath 'Coke/img/' (the results won't match with the groundtruth but with parameters from groundtruth_rect.txt trackers miss the can from the very beginning) All trackers cope with the task of can tracking.

2. DragonBaby
   Run with: python3 mean_shift.py --roi 160 83 100 100 --dpath 'DragonBaby/img/' As movements are fast CamShift window becomes large on the erty beginning and remains same throughut all sequence of images. Custom MeanShift goes out of window soon and the process interupts.

Note: Custom mean shift sometimes fails (centroid receives NaN coordinates; I wish I had time to investigate this). In this case program crashes with exception.

• for Coke dataset program doesn't crash when centroids are initialized as centroid = np.zeros(2)

But in this case meanshift converges much slower (up to 20 iterations) than it can.

• for DragonBaby dataset program doesn't crash when centroids are initialized as
  centroid = get_central_point(input_roi_box)

So we take the current central point of region of interest and meanshift coverges in up to 3 iterations on each step.