During exams, students leave their belongings on library seats to hold them and then disappear, so finding a free seat is hard. Occupancy sensors at every seat would be too expensive, but most study areas already have CCTV.
I led a team of three to build a proof-of-concept in Python. We recorded and labeled simulated surveillance footage of a study area from three angles, with four actors. To classify each seat from current and past frames, we combined COCO-pretrained person and chair detectors in TensorFlow with classical OpenCV techniques.
The system reached more than 90% soft accuracy at 10 frames per second in real time.