-
Notifications
You must be signed in to change notification settings - Fork 1
zhangjingling/Vision-Motion
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
Linxi Fan All the scripts run successfully in Matlab. =============== Challenge 1 =============== I use the function normxcorr2() to do the template matching. This can be much faster than a double for-loop with corr2(). It is very tricky to match the template close to the image boundary. My algorithm takes special care of these cases. On top of the brute-force search, I impose a few filtering rules to make the optical flow more robust. For example, any flow vector that has length greater than a certain (variable) threshold will be ruled out. This is why you see less noise (random arrows) in the area where there's very few texture (e.g. the black region at the bottom right). In this way, the annotated image looks much better. I experimented with a very dense field and saved the results to the folder "flow_result". The flow field looks highly smooth and robust. The screenshots are taken manually for better resolution. The script saves the annotated screen with the function 'saveAnnotatedImg' we used before. The output file names are "result_<i>_<i+1>.png". The actual grid density I set in runHw5.m is lowered for speed. It takes around 30 seconds to render 5 annotated images on my laptop. - Debug1a Parameters: search_half_window_size = 10; template_half_window_size = 4; grid_MN = [32, 24]; - Challenge1a search_half_window_size = 10; template_half_window_size = 15; grid_MN = [32, 24]; =============== Challenge 2 =============== The object tracker works really well on all three data sets. The annotated image sequences are saved to "*_result" folders. I modified the trackingTester() signature slightly. Now it takes a third boolean argument. If true, it uses imshow() to display the annotated image stack. If false, it silently saves the results to disk. The best number of histogram bins is found to be around 16 to 24. Note that the box I use is smaller than the bounding box that wraps the entire object. This makes the tracking much more robust. In the "basketball" data set, the tracer does not lose its target when the two green players overlap (frame 177).
About
Computer vision course: optical flow and moving object tracker
Resources
Stars
Watchers
Forks
Releases
No releases published
Packages 0
No packages published