Region-Based People Counting with Embedded System in Smart Building Environment

Smart building applications are growing that there are many kinds of implementations throughout the scope. This kind of application has a purpose to make convenience to humans without interrupting their activities. However, some research only focuses on improving object detection without considering the object's position. In this paper, the authors implement a program to count people based on designated regions in the camera's point of view of the building room. The architecture used in the paper uses RTSP-based IP cameras in each room connected to an embedded system. The embedded system uses the YOLOv4-tiny object detection model combined with KLT tracking. This algorithm is accelerated with Nvidia Deepstream, then outputs another RTSP link from the Deepstream. As a result, a single input can produce 23.2 frames per second (FPS), whereas two inputs can produce 9.3-10.7 FPS with FP16 precision. Two inputs can produce latency at 1–2 seconds with 10 FPS input framerate.