see this book:

Wolfgang Koch, Tracking and Sensor Data Fusion : Methodological Framework and Selected Applications, Springer Heidelberg, 2014.

Hi,

in a bayesian framework, you can solve your non-linear filtering problems by using the particle filtering. It is based on Sequential Monte Carlo methods and these kinds of filters can handle all situations (non-linearities in models, non-gaussian assumptions). You can find the basis of this filtering theory in the paper of Arulampalam or Doucet for example, which are used in some of my papers.

If your applications concern multiple object tracking, you can use the classic association methods (JPDA, MHT), or the recent PHD framework, avoiding all association errors.

Article Sequential Monte Carlo Methods in Practice by Arnaud Doucet;...

Conference Paper A Track-Before-Detect Approach for Extended Target Tracking ...

Article Book review: Bayesian multiple target tracking, 2nd edition

Conference Paper Visual Feature Tracking Based on PHD Filter for Vehicle Detection

The state of the art object tracking algorithm using particle filtering and compressed sensing for dimensionality reduction has been explained at the below given link:

https://www.researchgate.net/publication/261706562_Real_time_Visual_Tracking

Data Real time Visual Tracking

Particle filter or Monte-Carlo simulation method is for non-linear non-Guassian situation

There are many methods for dealing with non-linear/non-Gaussian estimation problems in tracking. An introduction to different tracking techniques can be found in my IEE 2005 survey paper, which covers the material prior to the growth of particle filters and FISST/PHD filtering work, which is now very popular. Gaussian mixtures is another popular framework, but you should always tailor the solution to the specific problem. EKF or iterated EKF, UKF, mixture model, particle filter, etc. They all have their advantages and disadvantages. A good idea is to implement two or more and see what works best for your problem and data. Bear in mind that PHD methods typically provide unlabelled tracks in the case of multiple objects. The other point I will make is that the non-linearity has to be fairly severe (compared with the sampling time) to justify the use of the more advanced methods over simple schemes like EKF. No point using a sledge hammer to crack a nut ;-)

Article Taxonomy of multiple target tracking methods