To provide safe autonomous driving to customers, the automotive industry faces a huge challenge to test and validate the self-driving functions. It is estimated that self-driving cars will cover about 240 million real or virtual kilometers on the journey to being mass production-ready. In other words, each function in the sense-plan-act loop of autonomous driving has to undergo a stringent test and validation process using a large amount of real and simulated data. Thus, the generation of reference data for test and validation purposes at an industrial scale is a crucial topic and requires a high degree of automation. In this paper, we propose a deep learning-based approach named RoadNet to generate reference road data for testing and validating the in-vehicle lane detection function. Furthermore, we present our processing pipeline running in a data center which fully automatizes the generation of reference data through RoadNet. We evaluate the accuracy and generalization ability of the proposed method using the automation pipeline. Although improving the quantitative evaluation results is still work in progress, we believe that our solution to productionize a deep neural network is of great importance for the cost-sensitive automotive industry.
Lidar-based Deep Neural Network for Reference Lane Generation
2020-10-19
3329688 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
LIDAR ENHANCED POLYNOMIAL GENERATION FOR LANE CENTERING
Europäisches Patentamt | 2023
|Europäisches Patentamt | 2023
|