I want to get some advice on what is the best method for me to implement a certain computer vision task.

My community college has a drone club, which I am a part of. We entered a competition where we have to build a drone that can autonomously deliver a payload. One of the requirements is that the drone have object detection capabilities. The drone must be able to see two types of objects:

1. A person
2. A "standard object" which is a normal sized piece of paper that has a colored shape and alphanumeric on it. The standard object is how the drone will know where to drop off its payload.

The possible shapes include: circle, semicircle, quarter circle, triangle, square, rectangle, trapezoid, pentagon, hexagon, heptagon, octagon, star, and cross. Valid colors include: white, black, gray, red, blue, green, yellow, purple, brown, and orange.

This whole thing is going to be running on a Raspberry Pi, possibly with a neural network accelerator attached if needed. The drone is going to be flying at least 15 mph and at least 75 feet above the ground, and it needs to see the piece of paper from up there.

I know a bit about how machine learning works, but that's it. Should I crack open a tensorflow textbook and try to build everything from scratch? Or is there some kind of premade general shape and color recognition that I could use and patch together so it recognizes what I want? Maybe I can use a pretrained model and adjust it to my purposes, and avoid having to learn all of tensorflow.

I also have a question about training. If at all possible, I want to avoid printing out all the possible combinations of the standard object (there are 140) and getting training data the normal way. Could I somehow generate a training set for the standard object detection? For example, by getting footage from the drone's flight and inserting images of the paper into the video?