Segmentation for efficient browsing of chronical video recorded by a wearable device

Video segmentation is often the first and a key step in many video analysis problems, such as in analyzing chronically recorded daily life video using a wearable device. Pre-segmentation is conducted based on shot boundary detection in multiple feature spaces. Subsequent boundary merging and refinement are based on shot duration thresholds and keyframe comparisons. Different from traditional keyframes, our key frames have statistical characteristics which best match those of the video segments they represent. Experiment results demonstrate that our method is effective in rapid video browsing.