A statistical approach to threshold selection in temporal video segmentation algorithms

Almost all temporal video segmentation algorithms employ thresholds. Although the selection of thresholds considerably affects the algorithm performance a framework to choose them effectively does not exist. This paper introduces a framework for selecting statistically optimal thresholds in temporal video segmentation algorithms. Statistics for various temporal video events, such as shot boundary, zoom and pan events, are collected. Then, optimal thresholds are estimated so as to minimize a statistical cost function defined in terms of the complied statistics. Results with several real video sequences containing over 150,000 frames show improvements ranging from 5%-20% in true detection and false alarm rates