Towards a low-cost embedded vision-based occupancy recognition system for energy management applications

This paper focuses on the development of a low-cost real-time occupancy detection system for people using convolutional neural networks. The proposed detector was implemented in an embedded system composed of a Raspberry Pi 3, an Intel neural computer stick accelerator, and a control circuit containing a relay, a transistor, and the Raspberry output ports. The model was calibrated by varying two parameters: intersection-over-union score and probability size, both necessary to achieve high level of confidence when detecting a person. An experiment was carried out as proof of concept of the system under different test scenarios such as walking fast with poor and optimal lighting conditions and strolling with good lighting. As result, the system obtained a confidence level above the 80% on all test scenarios.