MYA98IAPRWorkshoponMachine.PDFVIP

MYA 98 IAPR Workshop on Machine Vision Applications,Nov. 17-19. 1998. Makuhari, Chiba, Japan 12-2 Visual Feature Tracking with Automatic Motion Model Switching P. Tissainayagarn and D. Suter email: [prithi, suter]@zeno.eng .monash.edu.au Dept. of Electrical and Computer Systems Engineering Monash University, Clayton Campus, Victoria 3 168, Australia. Abstract of multi-target tracking. Cox et. al. later modified the In thispaper we present a motion determination and MI-IT with significant computational efficiency tracking technique based on the combination of (fig. 1). See [7,1] for complete details. Bayesian multiple hypothesis and a Multiple Model Filtering (MMF) algorithm. Corner features appearing in the initialframe of an image sequence were predicted in the subsequent frames using an extension of the multiple hypothesis algorithm (MHT [I]) based on different motion models. The collection of data provided by such a system was thenprovided to a Mh1F algorithm to determine the correct motion l ~ k b e r t h y p o m e ~ of features. We considered drferent order velociry and acceleration modelsfor the MMF algorithm and MdFM M g u i y hbtlix n applied them to two image sequences, the PUMA and t Uk) obwcd~ I U Toy car sequences. The study shows that the method proposed can distinguish between drfferent motions depicted in an ima