Analýza chodců pomocí obrazů

Abstract

This thesis focuses on the detection, prediction, and analysis of pedestrian behavior. In the context of autonomous vehicles, it is important for vehicles to correctly evaluate pedestrian behavior and predict their intentions in order to reduce the risk of collision. Various methods and models will be tested for person detection (YOLO, MediaPipe, OpenPose) as well as for the prediction of pedestrian movement, gaze, and gestures (LSTM, GRU, FFSTA, CNN, Kalman Filter). The methods and models will be trained and their functionality and accuracy will be compared on several datasets, including a custom dataset. Finally, a simple application will be developed, utilizing the best models from each prediction modul.