Mean shift and feature selectionOutlineMean shift: (1)historyMean shift: (2)IdeaSlide 5Mean shift: (3)CommentsMean shift: (4)feature selection comes up…Feature selection: (1)IntroductionFeature selection: (2) histogramFeature selection: (3) commentsExperiment demosSlide 12Conclusion and future work1Mean shift and feature selectionECE 738 course projectZhaozheng YinSpring 2005Note: Figures and ideas are copyrighted by original authors2OutlineDetecting, tracking and recognizing movingobjects in video sequences is a hot topic.•Mean shift algorithm---target localization•Feature selection---target representation•Experiment demos•Conclusion and future work3Mean shift: (1)historyY. Cheng. “Mean Shift, Mode Seeking, and Clustering” PAMI 1995D. Comaniciu, P Meer. “Robust analysis of feature spaces: color image segmentation” CVPR 1997 GR Bradski. “Computer vision face tracking for use in a perceptual user interface”. Intel Technology Jounal 1998. D Comaniciu, V Ramesh, P Meer. “Real-time tracking of non-rigid objects using mean shift”. CVPR 2000 Best paper award and patent filed.4Mean shift: (2)IdeaSsSsxsKsxsKxm)()()(..01)(woxi fXKxxmshiftmean )(:)(: xmxUpdate 5Mean shift: (2)Idea•Mean shift algorithm climbs the gradient of a probability distribution to find the nearest domain mode (peak)@R. Collins CVPR 2003 @Comaniciu PAMI 20036Mean shift: (3)CommentsInstead of exhaustive search in the window, the gradient information provided by the mean shift is used to reduce the time costMuch better than moving the search window pixel by pixel and scanning row by row7Mean shift: (4)feature selection comes up…•Given a likelihood image, locate the optimal location of the tracked object•The likelihood image is generated by computing, at each pixel, the probability that the pixel belongs to the object based on the distribution of the feature8Feature selection: (1)Introduction•Feature description approaches–Statistical descriptor–Structure descriptor–Spectral descriptore.g. intensity, color, texture, appearance, shape, motion, depth and so on.@Bradski Intel Technology Journal 98’9Feature selection: (2) histogram•Color histogram is widely used as object feature@Bradski Intel Technology Journal 98’Red: 1D cross section of an subsampled color probability distribution of a imageWhite: search windowBlue: mean shift point10Feature selection: (3) comments•Foreground and background appearance changes every frame, so the object features need to update•Color tracking affected by colored lighting, dim illumination or too much illumination•More object information should be used to increase the tracker robustness@Bradski Intel Technology Journal 98’11Experiment demosOriginal frameProbability distributioncolor tracking12Experiment demosNon-rigid shape change lighting change on objectObject with non-identical colorsimilar color with background13Conclusion and future work•Mean shift algorithm can use view-point insensitive appearance model (like color histogram) to track non-rigid object on changing background•Tracking success and failure will depend on the likelihood image which relates to the feature