In the project KI Data Tooling, a first successful measurement campaign for real data collection was carried out between the 17th and 19th May. In an interview, Manuel Hetzel (University of Applied Sciences Aschaffenburg, UAB), Jihad Miramo (Bosch) and Tobias Wagner (Valeo) reported on the extensive preparations, the practical implementation and what lessons they have learned for the next measurement campaigns.
Let's start with the motivation: What were the objectives of the measurement campaign in Aschaffenburg?
Manuel Hetzel: The primary aim was, of course, to collect an initial set of real world data for KI Data Tooling. In the project, this is usually done through two different methods: On the one hand, the test vehicles drive freely through Germany and collect data that way. On the other hand, we can also fall back on two research intersections equipped with additional sensors: one in Braunschweig, one in Aschaffenburg. Here we have the opportunity to plan and replay specific scenarios and thus collect precise data – exactly what we did during the measurement campaign.
Tobias Wagner: Another advantage of such a campaign is, of course, that we receive data on certain scenarios from three perspectives: Both test vehicles and the sensors installed around the research intersection record data of the same situation - so we can compare them afterwards and also learn a lot from the synchronization regarding space and time. These were the motives for carrying out this measurement campaign and, of course, for planning further measurement campaigns like this.
How exactly did you prepare beforehand?
Manuel Hetzel: Initially, the three project partners Bosch, Valeo and UAB were mainly involved. However, we naturally included all partners in the preparation. Each partner could fill out a kind of requirements list for certain scenarios or special requirements in a ticket system. Based on these tickets, we then created a schedule for the campaign. Of course, we were not yet able to respond to all requests, so for the time being, we proceeded to give everybody an equal share. Thus each partner gets some requirements fulfilled. We can then take the rest of the tickets into account in the coming measurement campaigns.
And how exactly did it work during the campaign?
Jihad Miramo: We divided the aforementioned schedule into time slots. For each time slot there were, beside the regular road users, two to three instructed actors – mainly students from UAB. Before each scenario, vehicles and extras were connected via a telephone conference so that the exact start could be timed and everyone moved along their predefined routes. When a manoeuvre was finished, the vehicles had to turn around again and everyone had to get back into the starting position for the next scenario. This way, we were able to record about 250 usable scenes during those three days at the research intersection. In most scenes both vehicles are involved so that we receive about 400 scenes in total from the vehicles’ point of view.
Tobias Wagner: In addition to the predefined scenarios, we were able to collect further data from the turning situations of the vehicles, so we received some bonus data there. This means that the Valeo vehicle for example was able to record about 300 scenarios.
Were there any unforeseen events during the measurement campaign?
Manuel Hetzel: There were no “typical“ unforeseen events. Of course, there were some road users - not our extras - who did not always adhere to the traffic regulations and, for example, ran a red light. However, these are pretty common occurrences in regular road traffic. In addition, we had some unusual scenarios in our script, e.g. pedestrians walking with crutches or pedestrians carrying a big ladder. In addition, the weather was also pretty varied - we had both sun and rain and were able to carry out good scenarios even at dusk.
What happens now with the data?
Jihad Miramo: At the moment, each involved partner is busy analysing their own data. We plan to upload all the data to a common platform by the end of the month to make it accessible to the other project partners. Then, we can also start the comparative analyses. Since this was first campaign of this kind, we are in the process of developing the tools and methods that are needed to perform such comparative analysis.
Were your expectations of the campaign fulfilled? What lessons did you learn?
Tobias Wagner: With regard to the data, we cannot yet draw conclusions prior to the evaluation - however, we are already very satisfied with the amount of data and the diversity of the scenarios. I think that in the end every partner will benefit from the collected data, which is a success.
Jihad Miramo: In addition to the data collected, we are also very satisfied with the organisation of the campaign, everything went well. Of course, there is always room for improvement in some areas - for example, planning more time for turning manoeuvres or for recharging the batteries of the vehicles during the next measurement campaign. However, these are typical lessons learned from such a campaign, which we will immediately implement in the next campaigns.
And finally: What is the significance of the measurement campaign for KI Data Tooling and for your further work?
Manuel Hetzel: Real data is one of the cornerstones of KI Data Tooling. Almost any of our work packages needs real data in one form or another – for example for comparing real and synthetic data, for the development of auto-labelling methods, for corner-case detection, for compression or for data augmentation. With the measurement campaign, we have therefore set an essential foundation for the further work of the project. That is why further measurement campaigns are already planned for the summer and November/December.
Jihad Miramo: I would like to add a non-technical significant benefit of this measurement campaign, namely that we were finally able to meet in person for the first time. This strengthens our cooperation in this project and further facilitates the upcoming campaigns.
Image: TH Aschaffenburg