Road User Detection and Analysis in Traffic Surveillance Videos

Road user data collection and behaviour analysis has been an active research topic in the last decade. Automated solutions can be achieved based on video analysis with computer vision techniques. In this thesis, we propose a method to estimate traffic objects’ locations with state-of-the-art vision features and learning models. Our focus is put on the applications of cyclist’s helmet recognition and 3D vehicle localization. With limited human labelling, we adopt a semi-supervised learning process: tri-training with views of shapes and motion flow for vehicle detection. Experiments are conducted in real-world traffic surveillance videos.