stem/Signal Proc/Image/Tracking.md

42 lines
1.2 KiB
Markdown
Raw Permalink Normal View History

# Challenges
- Clutter
- Distractors
- Occlusion
- Hidden by other objects
- Objects appearance can evolve
- Rotation, scale, camera viewpoint
- Take new template every n frames
- Take new template when confidence falls below threshold
# Background Tracking
- Static camera
- Capture clean shots of background
- Object present
- Average enough footage
- Background image - current video frame = difference image
- Threshold for binary mask
# Nearest Neighbour Tracking
- Decide component with closest centroid using previous centroid
- Not good for occlusion
- Will snap to next candidate
# Blob Tracking
- Build colour model of object
- Eigenmodel
- Mask of pixels that match object
- Use centroid as location over time
- Pick connected component with centroid closest to previous location
- Good for distinctive colours
- Not for practical situations though
# Template Tracking
- Sample distinctive patch from image
- Search all positions in video for patch
- Use cross-correlation
- Illumination changes
- Brightness is uniform shift of greyscale values up or down
- Correlated to the mean pixel value
- Subtract means in template and frame to give invariance
- Normalised cross-correlation