Využití mixtury gaussiánů pro detekci anomálií v časových řadách

Abstract

This thesis deals with the detection of anomalous walking behaviour using a combination of the MediaPipe library, a statistical Gaussian Mixture Model (GMM) and a Long Short-Term Memory (LSTM)neural network. The aim of the project is to create a system capable of analysis of a person’s motion captured by a camera and identify deviations from normal behaviour. MediaPipe is used to extract key points of the human skeleton from the video, and the extracted data is converted into vector form and processed in two ways. The first method uses GMM to classify normal behavior and detect anomalies based on a probability score- if an observation does not significantly match any of the existing clusters, it is considered anomalous. The second method uses an LSTM model that is trained on motion sequences to capture the temporal dependence in gait. The LSTM model can effectively detect unusual movement patterns over time that differ from the learned normal patterns. The result is a comprehensive tool for detecting anomalies in human motion that may find applications in security, healthcare, or sports analysis.