The difference seems to be that cmacs are using a symmetric encryption additional to the hash-function while hmacs process the key within the hash-function itself. Is that correct?
There is no "normal" collision-resistant hash involved with CMAC. See below for details. For HMAC, the internals of the hash function are exploited to process the key, however it can use the hash function as a black box.
HMAC was standardized in RFC 2104 and NIST FIPS 198-1. The idea of HMAC is to exploit the common Merkle-Damgård structure of hash functions. This is done by first invoking the compression function with the pre-processed key, then using that state to process the message normally. Finally, the output of the previous step is used as a message again in the same manner as the original message was in the previous step, though this time with a different derived step.
CMAC is defined in NIST SP 800-38B which has a nice figure explaining the mode:
Where the $\oplus$ denote XOR, CIPH is a block cipher, the MSB operation is truncation, $K,K1,K2$ are keys derived from the main input key via doubling in a polynomial field.
The message is divided in blocks of at most CIPH's block size.
Are there some situations when a cmac shall rather be used than a hmac?
Yes, the typical recommendation is that if you either have hardware acceleration for a cipher or if you want to minimize code size and provide other functionality relying on a block cipher in your code.