I'm trying to find a good object detection model to use in my application, running on the coral TPU, but have a few questions about where to find a good option.

My application is watching a security camera for "interesting" objects, and notifying me in real-time when detected. As such, I have the following requirements:

• fast. I would like to analyze images at a rate of around 5-10 frames per second, for quicker notification (you'd be surprised how far a car can move in one second)
• accurate. I don't want to be notified that there is a train in my driveway every time the shadows change :)

Of course, both of these are "soft" requirements, but ideally are the goal. So far for model selection, all I have found is the ones on the coral.ai main page: https://coral.ai/models/object-detection/ Which leads to my questions:

• Those models are listed as "not production-quality models". If that is the case, where might I find production quality models?
• If retraining is the answer to making them "production quality", how would I go about that? The instructions on that page give instructions for training the models to recognize additional object types, but I don't need that - I just need fast and accurate recognition of a handful of object types (people and the various forms of transportation they might use to arrive in my driveway, plus common mammals such as dogs and moose). Also, I would need to know where to get training materials (I could pull frames off my camera, but that would be a royal pain).
• Sticking to the models on that page, it looks like I have a choice of "fast", such as the SSD models, or "good", such as the EfficientDet-Lite models. Is that going to be generally true? I have also noticed the EfficientDet-Lite models use a LOT more CPU, even though they should be running on the Coral TPU.
• Are there other differences between the SSD models and the EfficientDet-Lite models that would recommend one over the other?