Moving Object Detection with Background Model based on spatio- Temporal Texture Ryo Yumiba, Masanori Miyoshi,Hirononbu Fujiyoshi WACV 2011
Outline Introduction ST-patch features Background subtraction – Generation of background model – Moving object detection – Update background model Experimental result
Introduction Background subtraction is a common method for detecting moving object. – Advantage: not requiring previous knowledge of moving object – Problem: cannot discriminate moving objects from background when these background change significantly Two approaches to generating BG model covering changes: – Pixel-wise background model – Patch-wise background model Proposed method cover global changes by using appearance information and it cover local changes by using motion information
ST-Patch Features appearance information motion information [7]
ST-Patch Features Patch size: 15*15*5(frame) – Appearance components differ between tree and road without regard to motion – Motion components increase according to temporal change – Motion components differ from transitions of sunlight and waving of tree
Generation of Background Model Use Gaussian mixture distribution of ST-patch to generate background model. Parameters are calculated previously from examples of background video using EM algorithm Background changes generally differ according to location calculate Parameters at each block
Detection of moving object
Update of Background Model It is difficult to generate a background model that wholly covers changes in background in advance update background model during moving object detection Means of distributions are close
Experimental Results
Experimental Result -- Outdoor Scene Waving tree, sunlight Use 1179 frames without pedestrians to generate background model Regard 1359 frames as frames with moving objects
Experimental Results -- Outdoor Scene 289 frames # FN ↓ ( ∴ background updating)
Experimental Results -- ceiling light scene
Experimental Results Outdoor Scene
Experimental Results