So after all the time I'm back with my feedback.
It didn't solve the accuracy in a way I wanted it to be, but it was getting a little bit better by changing the kernel of the census-trnasformation with: depth.initialConfig.CensusTransform.KernelSize.KERNEL_7x9
But I think I wont get the accuracy that I need by stereovision, because census transformation (or disparity calculation) isn't accurate enough, even if the images are very well detailed.
Nevertheless thanks for your help.