Models of Tracking by Decomposition of the State Space
Séverine Dubuisson · 2015
This chapter suggests three approaches exploiting the decomposition of the state space, all of which rely on the scheme based on partitioned sampling (PS). A first approach, dedicated to multi-object tracking, estimates conjointly the joint state of the objects and the order in which the objects need to be processed. This approach is shown to be competitive, even better than certain state of the art methods, such as PS, branched PS (BPS) and dynamic PS (DPS). The two others, used in the context of tracking one or several articulated objects, rely on a principle of simultaneous processing and the permutation of sub-samples defined in the independent sub-spaces and proving a first solution for the estimation and a second one for resampling. This idea of permutation makes it possible to, in particular, overcome a first barrier by estimating densities better than the other state-of-the-art approaches, while reducing the computation time.