Predicting Multiple Target Tracking Performance for Applications on Video Sequences
Machine Vision and Applications
This paper presents a framework to predict the performance of multiple target tracking (MTT) techniques. The framework is based on the mathematical descriptors of point processes, the probability generating functional (p.g.fl). It is shown that conceptually the p.g.fls of MTT techniques can be interpreted as a transform that can be marginalized to an expression that encodes all the information regarding the likelihood model as well as the underlying assumptions present in a given tracking technique. In order to use this approach for tracker performance prediction in video sequences, a framework that combines video quality assessment concepts and the marginalized transform is introduced. The multiple hypothesis tracker and Markov Chain Monte Carlo data association methods are used as test cases. We introduce their transforms and perform a numerical comparison to predict their performance under identical conditions.