On the road to the commercialisation of an automated vehicle, its software needs to be deeply tested. Due to the system’s automated nature, these tests are unique; however, if conducted on public roads and a system failure occurs, a catastrophic accident is highly likely. Therefore, to avoid this scenario, the system must be first tested in controlled environments, such as test tracks or simulated urban areas, where factors like traffic, weather, and pedestrian interactions are carefully managed. This kind of testing is quite demanding both in up and down time, and it cannot cover all the cases specified by the system constraint definitions, particularly those concerning actor safety. Simulation offers a solution to the limitations of a controlled environment.
Although simulated test environments require a significant initial setup, once they are running, tests can be executed automatically, and with significantly smaller downtimes. The compounding of errors forces the initial simulation setup to be done with a very high degree of accuracy, furthermore, additional details result in a smaller simulation-to-reality gap. This initial step counts with Digital Models from the test track, weather validation, and mainly sensor models and noise models development for these specific test cases.
Named after the action of practise a play or song for a later public performance, REHEARSE (adveRse wEatHEr datAset for sensoRy noiSe modEls) comes to help with the minimisation of this gap in the initial setup from the testing system. It is a dataset from data collected in the CARISSMA Outdoor Track in Ingolstadt, Germany, and the CEREMA Pavin fog and rain test chamber in Clermont-Ferrand, France. The usage of outdoor and indoor test sites allows for a head-to-head sensor data comparison between different environments. Furthermore, the test sites are complementary in their test capabilities, further strengthening the dataset.
In each of the test tracks, data is collected from two different cameras, a 4D RADAR, and two different types of LiDAR. The sensors and target are static. The targets measured are EURONcap-validated targets in the shape of a car, pedestrian, and cyclist. To further improve the dataset robustness, validated camera, RADAR, and LiDAR targets are also measured. With the focus on adverse weather, REHEARSE has data on rain, fog, snow, and clear weather conditions, in a wide intensity range, in day- and night-time. The adverse weather caused in the scene is validated using a broad selection of meteorological instruments, which allows a precise weather simulation. The dataset is also shipped to the end user using a standardised format to further facilitate its usage and make it more widespread.
The ROADVIEW project has already utilised REHEARSE, for the development of sensor noise models (seen here), the development of a 4D RADAR and LiDAR model for simulation, and its usage has even been further amplified by the novel point cloud annotation method developed by ROADVIEW partners.
REHEARSE provides invaluable data that makes it possible for a sensor noise model under harsh weather conditions to be developed, as well as a broad range of algorithms in target detection under harsh weather. It also provides a visual comparison between different types of sensors, the usage of more than one type of camera and LiDARs allows for future projects to compare the efficacy of the selected sensors in harsh weather.
