A solution can be
- Detecting presence or not --> background subtraction, since the place where an object will be is known
- The result can be cleaned using morpholocal operations and blob analysis, both in openCV.
- Pass the bounding box of the car to the ANPR interface
You could even optimize this and take only a ROI of the found blob, since you will know due to the fixed cam, where the license plate will be with some buffer.