Kalman filters require that observed variables have Gaussian distributions.
Frequently this isn't the case. Lots of real world systems have more complex distributions, like for example "which lane of a highway am I in given this GPS position" (it's likely you are in one of the lanes, but possible although unlikely I am between lanes).
Particle filters use more computation than kalman filters, but they can deal with arbitrary (and unknown) distributions of variables. Considering very few applications of this kind of system are limited by compute power, a particle filter is almost always better suited. (especially when you have a GPU available, which is a perfect fit for particle filters).
Frequently this isn't the case. Lots of real world systems have more complex distributions, like for example "which lane of a highway am I in given this GPS position" (it's likely you are in one of the lanes, but possible although unlikely I am between lanes).
Particle filters use more computation than kalman filters, but they can deal with arbitrary (and unknown) distributions of variables. Considering very few applications of this kind of system are limited by compute power, a particle filter is almost always better suited. (especially when you have a GPU available, which is a perfect fit for particle filters).