Dynamic Input Generation for the Development of Active Safety Perception Algorithms.

Enhanced perception algorithms are the key requirement for the introduction of more sophisticated active safety functionalities to urban areas. In order to realize the principles of test-driven development for such systems, either a set of representative and comprehensive test-drive logs needs to be available up front or a sufficiently universal simulation environment for virtual maneuvers should be employed. In this article a case study of developing a radar-based estimator for target heading is considered. This relatively standard problem serves as an illustrative example to assess the merits of an inherently simplified stimulus generated from virtual maneuvers and its limitation in comparison to real measurements. It is argued that a convenient modelling approach for a virtual scene simulation applied from an early development stage can further improve the quality and integrity of active safety projects, especially when they involve multiple sensor types. In reality the stimulus for vision, radar, lidar and ultrasonic sensors has a single origin: the actual scene around the ego-vehicle. It is therefore beneficial for a simulation environment to reflect such structure. The article also provides the background theory and approaches to the assessment of the considered perception algorithm from an early concept stage to in-vehicle tests.