Video Tracking G. Medioni, Q. Yu Edwin Lei Maria Pavlovskaia
Goal Track moving objects in a video stream
Linking frames Each frame registered with a satellite image
Detecting Moving Regions A sliding window with the center frame as the reference Register each frame in the window to the reference A region is moving if it differs from the registered frames Moving regions are grouped into tracklets
Tracklet Association Motion Target remains within a reasonable distance between frames Appearance Target has similar color distribution between frames
Tracklet Evolution Temporal moves Change labels Spatial moves Change rectangles at one instant
Merge
Pre-processing Goal Enhance given video before tracking Methods Auto levels Adaptive auto levels
Auto Levels
Histogram of pixel values
Auto Levels Modified histogram
Auto Levels in Video Concerns Algorithm should be fast Do not need to perform histogram computations for each frame Can not treat each channel separately
Adaptive auto levels
Minima cutoffsMaxima cutoffs
Adaptive Auto Levels
Post-processing Identify tracklets that are too short Highlight tracklets of interest Renumber tracklets Display tracklet labels
Highlighting Tracklets Identify tracklets that are too short Highlight tracklets of interest
Renumber tracklets
Displaying Tracklet Labels Goal: intelligently display a label next to every tracklet box
Displaying Tracklet Labels Desired specifications for label placement: – Label must be near corresponding box – Labels must be inside image boundary – Labels should not overlap – Labels should be far from other boxes – Labels should be far from box corners – Labels should not jump from frame to frame – Algorithm must be fast
Final Result
Thanks!