Detekce anomálií v chování řidiče s využitím obrazové analýzy

Abstract

This bachelor’s thesis focuses on detecting anomalous driver behavior from video data using neural networks. The goal was to design a method capable of distinguishing between normal and abnormal behavior based on temporal motion patterns. Key body landmarks were extracted from video frames and used to create different types of input features. An LSTM autoencoder was used to model normal behavior and was trained only on normal data. Anomalies were detected based on reconstruction error. Different feature representations, network architectures, and window lengths were compared.