You might also be interested in a document that @Frank mentioned in this discussion: We have an application note that describes how to interface 
to OpenCV. This is not directly CUDA, but the concept is the same: If you want to use the content of a image as the input buffer for a CUDA kernel you’ll effectively need to do the same things that you’d need to do to use a image as the data source for OpenCV (and vice versa). You’ll probably make things a lot easier if the image width is a multiple of four, otherwise all you need to verify is that GetLinearAccess pointed out by @Sebastian returns true (which is usually the case unless you made use of e.g. the RotateImage flag in the GenICam.ini file.
You can download the full document from here: https://ftp.commonvisionblox.com/webdavs/forum/documents/CVBInterop%201.1.0.pdf
@Sebastian: It is not apparent from this post, but @pavik1994 is using a TX1 which means GPUprocessing + HLSL is not an option.