A Normalized Mutual Information Estimator Compensating Variance Fluctuations for Motion Detection

In motion event detection, an information measure based score function is sensitive to variance fluctuations between sample intervals. In this work, a score function based on a normalization of mutual information measure is proposed to tackle the problem of variance fluctuations. The mutual information is normalized by the maximum entropy which is related to the sample variances in comparison. An estimator using the normalized mutual information measure is implemented by neural networks with random setting of hidden neuron parameters. This estimator is tested by change point detection and motion event detection in experiments. Experimental results show that the normalization scheme in the estimator improves the sensitivity of event detection.