Entropy-based Approach for Detecting Feature Reliability

Although a tremendous effort has been made to perform a reliable analysis of images and videos in the past fifty years, the reality is that one cannot rely in 100% on the analysis results. In this paper, rather than proposing yet another improvement in video analysis techniques, we discuss entropy-based monitoring of features reliability. Major assumption of our approach is that the noise, adversingly affecting the observed feature values, is Gaussian and that the distribution of noised feature approaches the normal distribution with higher magnitudes of noise. In this paper, we consider one-dimensional features and compare two different ways of differential entropy estimation—histogram based and the parametric based estimation. We demonstrate that the parametric approach is superior and applicable to identify time intervals of a test video where the observed features are not reliable for motion detection tasks.