Ground truth annotations are established through manual scene analysis, where key situations (spatial relations, weather/road context, and actions/events) are identified and labelled.