Stochastic Tracking in Image Sequences

Tracking features along image sequences is a Computer Vision problem which has evolved considerably in the last years. In fact, improvements have been made to try overcome difficult and ambiguous situations generated by cluttered backgrounds, occlusions, large geometric deformations, illumination variation or noisy data [1, 2]. On the other hand, with better computational and imaging resources, and the considerable research work done in the last decades, results are expected to be more accurate in the numerous applications of movement tracking, such as surveillance, analysis of objects’ deformation, medical image analysis and traffic monitoring [2, 3]. Usually, to track an object along an image sequence several measurements are made upon each image. However, those measurements do not always provide true values due to some of the above mentioned problems. The stochastic framework used in this work filters the imprecise measurements and maintains an estimation of the current state of an object. For that we have used the Kalman Filter and the more recent Unscented Kalman Filter [4]. To initialize such filters several assumptions should be made, but as results are obtained, the initial suppositions may be improved. In particular the dynamic model of the objects’ movement may be learned my means of hidden Markov models [5], and used to obtain more robust and precise estimates. In this work we will analyse the results obtained with the mentioned filters, as well as evaluate the influence of the used dynamic model. We will show that the proposed combination of approaches allows the successful tracking of complex motions with abrupt changes.