A recent rise of windshield mounted cameras in cars has given access to large amount of videos from real world traffic accidents. This paper aims to analyze what can be learned from that video data using current state-of-the-art computer vision techniques. We also discuss what is currently missing to fully understand these videos. A substantial number of accident sequences has been acquired and manually evaluated for this analysis. The ability to fully automatically analyze large numbers of traffic accidents from videos will allow to identify potentially dangerous traffic scenarios, improve driving safety and driver warning systems.