Performance Evaluation of Joint Tracking and Classification

Joint tracking and classification (JTC) is gaining momentum in recent years. Many algorithms have been proposed. However, not enough attention has been paid to JTC evaluation, although it is important in practice. In this paper, we deal with evaluating the goodness and credibility of JTC. For the JTC goodness, tracking and classification so far have been largely evaluated separately without considering the interdependence of tracking and classification. We propose a joint measure—joint probability divergence (JPD)—to quantify tracking error, misclassification and their interdependence. The basic idea of JPD is to measure the closeness between the cumulative distribution functions of the perfect JTC and the JTC to be evaluated. The proposed method has a number of attractive properties. Some results from algorithms can be regarded as self-assessments. The credibility problem is concerned with whether these assessments are credible or how credible they are. We define the credibility problem for decision and propose an associated noncredibility index (NCI). We also propose a joint NCI (JNCI) to quantify the noncredibility of estimation and decision jointly. Four examples are presented to demonstrate how well the JPD and JNCI reflect the joint performance of tracking and classification.