Transformation of adaptive thresholds by significance invariance for change detection

The detection of changes in image sequences often is the first essential step to video analysis, e.g. for the detection, classification and tracking of moving objects. As a binary classification problem, change detection is afflicted by the trade-off between two class error probabilities, viz. the rates of false positives and false negatives. In this contribution, we derive an adaptive two-threshold scheme to improve on this trade-off. The threshold selection for each pixel in the current frame is controlled by the previous detection result for this pixel. Since the test statistics are calculated from samples comprising several pixels within a local sliding window, a transformation of the thresholds from the single-pixel observations to decisions based on larger samples is required. Based on the fact that we can only model the null hypothesis, i.e., absence of motion, realistically, we suggest to transform the threshold under the constraint of a constant false-positive rate, or significance invariance. The resulting detection algorithm is only marginally more complex than a straightforward global thresholding procedure, while providing visibly improved results.