Data-driven Derivation of Requirements for a Lidar Sensor Model

Safety assurance in virtual driving simulation environments requires accurate sensor models. However, generally accepted quality criteria for sensor models do not yet exist. In this work, we investigate the model quality needed for a Lidar sensor model for virtual validation. We seek to answer the question, whether neglecting sensor effects in a simplified sensor model might lead to a measurable difference in performance of the sensor model compared to a real sensor. A data-driven approach has been chosen to identify relevant features for object classification in Lidar pointclouds which need to be accurately represented in simulations. The contribution of our work is two-fold: Firstly, we identify important features for object detection in point clouds from Lidar data. For this, we apply object classification algorithms to pointcloud segments, for which a variety of geometric, stochastic, and sensor-specific features have been calculated. Using filter models, principal component analysis (PCA), and embedded models, each feature is assessed and ranked on an individual basis. Secondly, we derive implications for Lidar sensor models based on our findings. We investigate variations in classification quality by succesively removing groups of features from our feature set. Our results show, that to make sensor models suitable for the validation of object detection algorithms, the accurate representation of simple geometric features in synthetic pointclouds is sufficient in many cases. Our method can also be used to support the derivation of requirements and validation criteria for sensor models.