Sledování pohybu osob ve vymezeném prostoru

Abstract

This thesis deals with observations of people in a defined area. The main goal is their re-identification, in other words identifying a previously observed person. This method assumes, that the given person did not change their clothes during the observation, which means that it's mostly suitable in a short time period. The OpenPose library and convolutional neural network were used to successfully accomplish this task. The MARS dataset, which is tailored for a re-identification of people, was used with slight modifications for training of the neural network. This paper also describes a designed method for extracting features of people from given frame. These features were used to build a matrix of size 36 x 20 pixels and was used as an input to the convolutional neural network. Design and training of convolutional neural networks is greatly described in this paper, as well as what methods and conditions are necessary to choose the best possible type of architecture. The output of the convolutional neural network is a descriptor of given person, which the system uses to find the closest match and predict the identity of a person in the frame. The accuracy achieved in this solution was 84.02% using the testing set.