Analysis of Environmental Influences for Simulation of Active Perception Sensors

Automated vehicles have an inherent safety risk for passengers and other traffic participant. Rigorous testing and safeguarding is needed for approving their operation on public roads. But testing in the real world is not only very time-consuming and expensive but also quite dangerous to participants and the engineers. Therefore more and more tests are relocated to the virtual world before they are performed on proving grounds and eventually in real traffic. In the real world, perception sensors of automated vehicles are subjected to a variety of adverse environmental conditions, such as fog, rain, snow, glaring sun light or road spray from other vehicles. As previous research already showed a severe impact on perception sensors, especially lidar, these influences need to be accurately represented in the virtual world and in models of the perception sensors. To systematically quantify the influence of the named conditions, they are first sorted into two main categories of object independent conditions, such as fog, rain, snow, and object dependent conditions, like wet pavement and road spray. For the first category, measurements in a stationary setup are recorded over a period of six months. Multiple lidar sensors with additional reference sensors for rain rate, temperature, sun brightness, visibility etc. deliver the data in this experiment. The measurements are sorted according to the weather condition and lidar values like the number of detections in the atmosphere are aggregated. This yields expectation values with respect to quantified environmental conditions. As a prominent example for object dependent conditions, road spray is examined in a second experimental setup. Measurements are taken with objects driving over artificially watered pavement on a proving ground. Water film and object velocities are systematically varied between experiment repetitions. The most prominent phenomenon in the recorded data is clustering of detections in the spray plume due to the turbulent nature of the spray. The clustering as well as detection probabilities within these clusters are used as expectation values for modeling. The gathered expectation values are then utilized to develop stochastic simulation models. The models are integrated into an lidar base model in a modular approach compliant to the Open Simulation Interface. The two main modeling approaches are adding false-positive atmospheric detections and attenuating or removing detections generated by the base model. Finally, the gained experiences from the measurements and model development are used to derive requirements for ground truth data quantifying the environmental conditions. The specified ground truth data serves as input to the simulation models