Protecting Vulnerable Road Users: Semantic Video Analysis for Accident Prediction