In this paper we present a probabilistic framework for tracking objects based on local dynamic segmentation. We view the segmentation to be a Markov labeling process and abstract it as a MAP problem. In the Bayesian formulation, we exploit the feature-spatial-measure distribution of local area as the conditional distribution. The feature-spatial vector is used to constrain the appearance of region while the measure vector is used to constrain the label of the pixels in the region. One drive force to the introduction of FSM distribution is the HMMF model that makes it possible to estimate the measure field by the minimization of a differentiable function. Mean-shift procedure and IFGT technique are used to further alleviate the computational costs. Very promising experimental results on synthetic and natural sequences are presented to illustrate the performance of the presented algorithm.
Joint feature-spatial-measure space: a new approach to highly efficient probabilistic object tracking
2005-01-01
312744 byte
Conference paper
Electronic Resource
English
British Library Conference Proceedings | 2005
|Probabilistic tracking in joint feature-spatial spaces
IEEE | 2003
|Probabilistic Tracking in Joint Feature-Spatial Spaces
British Library Conference Proceedings | 2003
|SET JOINT PROBABILISTIC DATA ASSOCIATION FOR RELATIVE SPACE OBJECT TRACKING
British Library Conference Proceedings | 2016
|Region Tracking via HMMF in Joint Feature-Spatial Space
IEEE | 2005
|