Object detection systems are embedded in many safety critical systems, such as autonomous driving systems. Assessing and minimizing the risks they introduce in these situations is a key concern. A universal definition of risk for all systems does not exist. Additionally, for a given system, the risk may change because it depends on the requirements, which evolve during development or production. This often results in multiple metrics being used, such as overall accuracy combined with the misclassification rate for a specified class. Existing research defines risk metrics for specific cases. However, the metric’s meaning depends on the context. We propose a domain-specific language (DSL) to address these issues, which can handle various metrics and contexts of object detection. Specifically, the DSL is used to define risks in object detection systems and to select datasets, which can improve a model’s performance by considering risk.