Zoom Detection in Video Sequences

Abstract

The work presented in this thesis is mainly based on the discriminative method. Zooming is one of the type of camera motion, it may be global or local. The aim of this thesis is to detect global zoom in the video sequences. Fisher linear discriminant (FLD), which works on the principal of linear discriminant analysis, has been used as the classifier for zoom motion classification in video sequences. The basic motivation behind the choice of FLD is that it mitigates the curse of dimensionality by using projection based approach and reduces the overhead of handling higher dimensional data by projecting it to lower dimension. To counter the outliers due to movement of objects as well as noise due to the encoding strategy which can be present at any location in the motion vector field in the video sequences, a quadrant based strategy has been employed. MLE (maximum likelihood estimator) is used to estimate the statistical parameters like mean and standard deviation of the orientation of MV’s and a novel weighting strategy has been used to counter the skipped macroblocks present in interframe motion field. Along with that mean deviation is used in place of mean which is less sensitive to noise. Scatter plot analysis and mathematical analysis have been carried out for feature selection. Features have been exploited from the orientation of the block motion vectors which renders this method computational advantages. Moreover, with significant maturity of coding standards like H.264 and MPEG-4, video are directly being stored in the compressed formats thereby opening up the area of the compressed domain video analysis. An extension of the proposed method has been carried out for zoom type classification i.e. for zoom-in and zoom-out classification from the identified zoom frames of video. The proposed method may find wide applications in the field of image and/or video processing.