The paper proposes a method for predicting when a person enters a forbidden zone during his trajectory following a video stream, considering individual body parts. The authors used the PP-TinyPose PaddleHub neural network model with its implementation based on two deep neural networks to detect key points of the human body. The paper considers an example of human position prediction from a continuous video stream in indoor trajectory tracking. The authors predicted each key point in the coordinate space of the video stream using a recurrent deep neural network algorithm.