Unsupervised Feature Selection for Detection Using Mutual Information Thresholding

This paper proposes a method for unsupervised selection of features for detecting important events in a surveillance context. While traditional feature selection requires manually annotated ground truth to choose the best features, we examine the possibility of exploiting the redundancy between a pair of independent data sources for selecting good detection features. Building on our prior work on mutual information thresholding, we show that strong agreement between data sources indicates strong detection performance. Experimental tests, combining visual and audio data, show that the best performing features can be automatically selected by taking advantage of the common information shared by the sensors.