This paper proposes a convolution particle filtering approach for extended object tracking. Convolution particle filters (CPFs) are likelihood free filters. They are based on convolution kernel probability density representation. They use kernels
to approximate the likelihood of the observations and represent the likelihood when it is analytically untractable or when the observation noise it too small. Hence, the CPFs represent a sub-family of particle filters with improved efficiency in state
estimation of nonlinear dynamic systems. A CPF is designed and implemented for track maintenance of an object with an elliptical shape. The object kinematics and its extent are estimated in the presence of dense clutter. This nonparametric filter is validated with a Poisson model for the measurements, originating from the target and clutter. Simulation examples illustrate the filter performance. It is shown that the CPF yields correct estimates of the joint probability density function of the state variables and unknown static parameters. The results obtained for the extended objects show that the CPFs provides accurate on-line tracking, with satisfactory estimation of the target shape and volume.