Object segmentation and tracking is such a natural and 'automatic' part of our visual perception that it's difficult to intuit how challenging it is to do with software.

It takes freshly deployed humans a while to master, and only then with fairly high bandwidth training data, so I wouldn’t feel too bad about the complexity of implementing it in software.