An event based approach to video analysis and keyframe selection

We propose an event based approach for locating keyframes in natural video through detection of locally correlated spectral targets. Temporal Decomposition (TD) is used to describe a set of spectral parameters of the video as a linear combination of a set of temporally overlapping event functions. This process provides preliminary information about keyframes, by selecting the frames located at event centroids as the keyframe candidates. No shot or shot cluster boundary detection is needed and keyframes are extracted directly from among event centroids that are much smaller in number than the number of frames. Generalized Gaussian Density (GGD) parameters, extracted from 2D wavelet transform subbands of the frames, are used as the spectral parameters in the event detection process and Kullback-Leibler distance (KLD) is employed as a measure to select salient keyframes. Experimental results confirm superiority of the proposed scheme over the conventional keyframe selection approaches.