The high cost of equipment that solves the problem of tracking a position of a person and its direction in real time is one of the factors that affect on the speed of development of the augmented reality industry negatively. The urgency of this problem is a background for the development of an affordable tracking system. This paper describes the software and hardware architecture of the system that solves the problem of three-dimensional tracking and pose classification in a confined space using neural network models. The uniqueness of this system lies in feasibility on strictly limited computing power and in absence of any sensors placed on the monitored objects. After setting boundaries of the active area, all necessary input data are provided by a static camera without an infrared filter. As an example of implementing a resource-limited solution, we present the assembly system based on a 3rd version single-board Raspberry Pi computer connected to the 2nd version Intel Neural Stick co-processor and the 2nd