3/6/2015 PortoICIAR’20041 Adaptive Methods for Motion Characterization and Segmentation of MPEG Compressed Frame Sequences C. Doulaverakis, S. Vagionitis, M. Zervakis, E. Petrakis Technical University of Crete (TUC) Chania Crete Greece
3/6/2015 PortoICIAR’20042 Problem Definition Video segmentation Abrupt (cuts) & Gradual transitions Zoom, Pan/Tilt Contribution: segmentation and identification of camera effects Processing on MPEG video Partially decompressed block data DC intensity approximation for blocks Coherent motion vectors for I, P, B frames
3/6/2015 PortoICIAR’20043 Twin Comparison (TC) Shot boundaries peaks on histogram differences Thresholds: Ta, Tb Ta = μ + α σ Τb = bμ Requires pre- processing,does not adapt to signal Τα Tb
3/6/2015 PortoICIAR’20044 Sliding Window (SW) Processing over W frames One Threshold: Ta(i) = μ(i) + α σ(i) Cut: < 5 frames Gradual: > 5 frames No preprocessing, adapts to signal Ta
3/6/2015 PortoICIAR’20045 Adaptive Method (AM) Ta(i) = μ(i) + α σ(i) μ(i) =μ(i-1)-c(μ(i-1)–D(i)) σ(i) = |μ(i) 2 – λ(i)| 1/2 λ(i)=λ(i-1)–c (λ(i-1)–D(i)) 2 c=0.05 Ta is computed at each i and depends on previous values No preprocessing, adapts to signal Ta
3/6/2015 PortoICIAR’20046 Direction Histogram Histogram of angles of motion vectors 8 angles multiples of π/4 for moving vectors Plus 0-th value for static vectors |v| < 1 static moving
3/6/2015 PortoICIAR’20047 Motion Characterization Analysis of variance σ motion histogram Normalized by number of intracoded vectors Zooming: the vectors are spread uniformely (max σ) Panning-Tilting: the vectors are concentrated at a single bin (min σ) Static camera: the vectors are concentrated at bin 0
3/6/2015 PortoICIAR’20048 Example pan pan zoom static camera Ta
3/6/2015 PortoICIAR’20049 Video Segmentation Method Cuts: the number of intracoded vectors in frame exceeds threshold Gradual transitions: combines motion and intensity information Difference of intensity histogram exceeds threshold Magnitude of motion vectors exceed threshold
3/6/2015 PortoICIAR’ Experiments Measurements over 17 videos Competitive methods correspond to thresholding by TC, SW, AM Each method is represented by its precision/recall curve as a function of the threshold parameter a
3/6/2015 PortoICIAR’ Abrupt Transitions (cuts)
3/6/2015 PortoICIAR’ Gradual Transitions
3/6/2015 PortoICIAR’ Future Work More accurate threshold estimation SW or AM gets trapped in local minima Detection of Cuts is fairly stable More elaborate methods for detection of gradual transitions and for cleaning- up false positives due to camera effects
3/6/2015 PortoICIAR’ Zoom Detection α=4.5 α=2
3/6/2015 PortoICIAR’ Pan/Tilt Detection α=2 α=4.5