andy
5a94c5ff1a
Affected files: STEM/AI/Kalman Filter.md STEM/Signal Proc/Convolution.md STEM/Signal Proc/Image/Tracking.md STEM/Signal Proc/Pole-Zero.md STEM/Signal Proc/Transfer Function.md STEM/Speech/Linguistics/Consonants.md STEM/Speech/Linguistics/Linguistics.md STEM/Speech/Linguistics/README.md STEM/Speech/Linguistics/Terms.md STEM/Speech/Linguistics/Vowels.md STEM/Speech/NLP/Jargon.md STEM/Speech/NLP/NLP.md STEM/Speech/NLP/README.md STEM/Speech/NLP/Recognition.md STEM/Speech/Perception/Perception.md STEM/Speech/Perception/README.md STEM/Speech/Speech Processing/Applications.md STEM/Speech/Speech Processing/README.md STEM/Speech/Speech Processing/Source-Filter.md STEM/Speech/Speech Processing/Vocal Tract.md STEM/img/english-phoneme-table.png STEM/img/formant.png STEM/img/pole-zero-attenuation.png STEM/img/pole-zero-feedback.png STEM/img/pole-zero-stable.png STEM/img/roc-right-left.png STEM/img/roc-two-sided.png STEM/img/spectrum-vocal-tract.png STEM/img/transfer-stable-unstable.png STEM/img/vowel-chart.png STEM/img/vowel-spaces.png
1.2 KiB
1.2 KiB
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