A closer look on traffic light detection evaluation metrics

Recognition of traffic lights is of crucial interest for intelligent vehicles. For camera based detection systems traffic lights constitute important objects that have some specialties in contrast to other traffic related objects: For a better sight of road users on traffic lights in many cases they are placed redundantly. Furthermore, at bigger intersections a particular traffic light often only denotes information for a particular lane. For a road user or intelligent vehicle only a relevant traffic light showing the information for the desired turning direction has to be found. When evaluating traffic light recognition systems for different use cases a different level of information needs to be evaluated as well. This paper provides a closer look on evaluation metrics and uses a real world dataset for providing examples of different metrics.