Collecting automotive sensor data for AI-applications is time and cost expensive. One solution is the usage of simulated data. However, the quality of the results highly depends on how close the simulated scenery is to reality. One main challenge is the definition of the properties of dynamic objects. A simulation of their behaviour over time is possible. But for this, extra software e.g. SUMO or CARLA is needed. This paper therefore gives a proposal without further simulation software. The proposal uses statistical representation from real automotive datasets. This gives the advantage of direct object placements. In addition the objects properties are as near as possible to reality. Another advantage is the creation of more data frames as a dataset can provide. First, several parameters are collected from real datasets. Then the parameters will be analysed regarding their statistical appearance. This analysis is then the basis for mathematical representations. Further metadata will be collected separately. Here, standards as openDrive are important tools. For validation, the objects are placed in a simulation environment (CARLA). To define the DOE here, sampling methods as Latin Hypercube Sampling are fed with the previously determined analysis. In the end, the synthetic data is validated by selecting the most similar frame from the real dataset.
Session:
Autonomous Driving I
|
| 13:40 - 14:10