What AI concepts, topologies1, algorithms, or SaaS can be used to recognize a person eating a chocolate.
For this question, image recognition draws from a real time feed, validating each of these steps in sequence:
- Using a camera app, the user begins recording.
- The user focuses on the subject's face, at which time the app attempts to recognize the person2
- Person holds a single piece of chocolate (i.e M&Ms) to the camera lens, at which time the app attempts to recognize it.
- The user puts the chocolate into their mouth, chews, and swallows, at which time the app attempts to recognize that AS AN ACTION.
- The app gives a completion message indicating success or failure.
I understand that we can use real time recognition for each step, but I don't know if there are concepts proposed or tested to validate the scene as a sequence or any of the three recognition steps individually.
The app should invalidate the scene if the subject is swapped with another subject, if the subject does not swallow the chocolate, or there is some other deviation from the expected sequence above.
 By topology in this context is meant the standard mathematical meaning of the term applied to the higher level connections that are likely needed to recognize sequences of actions. In this sense, the use of the term topology is not at the level dimensions of neuron layers or convolution kernels. Since process topology is normally considered prior to considering library dependencies or deployment concerns, the term topology is more appropriate than architecture in this question. (First things first.)
 I've already identified SaaS options for facial recognition.