Back to ASAM International Conference 2019

 

 

An Efficient Approach to Generate Tagged Traffic Scenario Models from Naturalistic Driving Data 

 

Dr. Xianpeng Lang        
Director of Autonomous Driving
CHJ Automotive (China)

 

With rapid development of autonomous driving, the need for tremendous traffic scenario data has never been greater, especially in algorithm research and Simulation. In recent years, extracting scenarios from naturalistic driving data (NDD) is recognized as a promising way to generate demanded scenarios. However, most existing approaches typically require a considerable manual work which made them impracticable. Furthermore, the extracted scenario data should be formatted and managed for according applications. In this paper we proposed a novel approach to efficiently generate traffic scenario models from NDD. Highly automatic processing is the novelty of the proposed approach. Firstly, we develop an event-aware data collection system, which use in-vehicle computer to monitor chassis bus signals and record raw data of demanded scenarios. Secondly, we create the static-Map around this area with the help of ego-localization results and Highly Defined (HD) Map. Thirdly, we apply deep learning algorithm to find out all traffic-participants and their relative locations from sensor data. Finally, we format environment-models with OpenSCENARIO Standard, as well as various tags for query usages. We also evaluate the difference between real world and generated models. Experimental results demonstrate that the average error of entities per model is less than 20cm (within 50m) or less than 50cm (50m–100m).