This manuscript deals with the problem of 3D object tracking in a multisensor framework. The object is here described by a CAD model. It avoids any image preprocessing that leads, generally, to loss of information. We develop a particle filtering method [6] that we call "correlation-based particle filter" (CBPF) to solve this non-linear estimation problem. The new proposed approach is applied to synthetic and real image sequences of complex 3D moving objects. The originality of this work consists of developing a centralized fusion method that uses, in an optimal way, the measurements delivered by the sensors. In order to optimally using the sensor outcomes, a centralized fusion approach is proposed. The method can jointly estimate 3D pose/motion parameters and track the object in the 3D domain, while many works have been developed in the image plane. Finally, we should mention that the method is not limited in terms of object structure and motion.

We present a method to track and estimate the motion of a 3D object with a monocular image sequence.

This paper details a model-based method to track object in an image sequence and jointly estimates its motion. It uses the grey-level images and no preprocessing, as feature extraction for example, is done on the sequence. The main benefits of this method which avoids this extraction stage is to allow a better localization of the shape. This estimation problem is modelled by state equations which describe the evolution of the object in the sequence and the related measurements and is solved by a non-linear filtering method: the particle filtering.

No abstract

No abstract