Week 4 Emily Hand UNR Basic Tracking Framework
Week 4 Emily Hand UNR
Basic Tracking Framework Template Tracking – Manually Select Template – Correlation tracking Densely scan frame and compute histograms. – 100 negative samples and 1 positive sample – The SVM classifier is updated with each frame. (Lib. SVM) Basic Idea – Tracker and Detector are independent
SVM Densely scan the neighborhood – Create Score Map – Determine location of object in frame Retrain SVM – Positive Examples • All previous templates – Negative Samples • Top 100 false positives • Create a score map from the entire frame
Some Results
Some Results
TLD Tracker Implementation in Open. CV/Matlab PN Learning – Lucas Kanade Tracker is used • Returns a Confidence – Trajectory correct if confidence>80% – P-constraints: all patches close to validated trajectory have positive label – N-constraints: all patches in surrounding of validated trajectory have negative label – These samples are used to update the detector unless there is a strong detection far away from the track
Problems Occlusion – Handled pretty well Appearance Changes (What we want to work on) – Initial Idea • Use background and non-targets as negative samples • When tracker fails due to appearance change, the target will be a non-negative sample in the neighborhood
Next Week Adaboost – Implement the Adaboost detector for our template tracking system Explore initial idea – Test out different methods for dealing with appearance changes
- Slides: 8