Low complexity event detection for video programs
A video processing method for detecting significant events from a video program includes computing short-time sub-band energies in the audio for plural audio sub-bands, detecting scene boundaries where a weighted sum of these short-time sub-band energies are less energy threshold for longer than an time interval, segmenting the video program into a plurality of scenes by the boundaries, removing scenes shorter than a segment time interval and classifying and ranking the remaining scenes by audio. A second segmenting and removal is based upon a second energy threshold and a second time interval or when energy in a lowest frequency sub-band is greater than a predetermined bass energy threshold. The first segment time interval may be recomputed based upon the distribution of length of the remaining scenes.
1. A video processing method for detecting significant events from a video program comprising the steps of:
computing short-time sub-band energies in an audio of the video program for plural audio sub-bands;
detecting scene boundaries from the audio of the video program by detecting whether the short-time sub-band energies are less than a predetermined first energy threshold for longer than a predetermined first energy time interval;
segmenting the video program into a plurality of scenes using the detected scene boundaries as boundaries;
removing scenes from the video program having a length shorter than a first segment time interval;
following said step of removing scenes detecting second scene boundaries in the audio of the video program by detecting
whether the short-time sub-band energies are less than a predetermined second energy threshold for longer than a predetermined second energy time interval, and
whether the short-time sub-band energy in a lowest frequency sub-band is greater than a predetermined bass energy threshold;
further segmenting the video program into a plurality of scenes using the detected second scene boundaries as boundaries;
further removing scenes from the video program having a length shorter than a predetermined second segment time interval; and
classifying and ranking remaining scenes based upon short-time sub-band energy in the audio from largest to smallest.
2. The method of claim 1 , further comprising:
following said step of removing scenes:
recomputing the first segment time interval based upon a distribution of length of the remaining scenes; and
further removing scenes from the video program having a length shorter than the recomputed first segment time interval.