The Critical Scenario Extraction And Identification For ICV Testing

Digital Twin(DT)technology extracts test scenarios from real-world driving data and is considered an effective solution for road testing of Intelligent Connected Vehicles(ICVs).How to extract and identify critical test scenarios from real driving data has become a research hotspot.Focusing on this issue,this thesis proposes a new DT test scenario selection method.Taking the collision risk and traffic factors into consideration,this thesis defines the criticality evaluation factors of such three typical application scenarios as Forwarding Collision Warning(FCW),Lane Change Warning(LCW),and Intersection Collision Warning(ICW).LSTM-AE-AM,a deep learning model for critical scenario recognition based on Long Short-Term Memory(LSTM),Autoencoder(AE)and Attention Model(AM),is thus proposed.The main work of this thesis is as follows.First,the collision risk and traffic factors of three typical application scenarios are analyzed,and the criticality evaluation factors of the scenarios are defined to support the research on the identification of critical scenarios.Then,the raw data set provided by Chongqing Vehicle Test and Research Institute is preprocessed to reconstruct vehicle trajectory in order to solve the problems of missing frames,abnormal points,and noise.Three typical application scenarios are extracted from the original data set to form a new scenario data set,which is clustered and marked.Following that,due to unbalanced samples in the newly constructed scenario dataset,features are constructed according to the spatiotemporal characteristics of scenario data,and DCGAN is used to expand critical scenario samples for data enhancement.A LSTM-AE-AM critical scenario recognition model is then proposed.The model uses LSTM-AE to extract high-dimensional data features for pre-training and then combines the attention mechanism to give more weight to important features,achieving the goal of critical test scenarios mining.At the concluding stage,comparative experiments are carried out on the dataset to verify the effectiveness of the data preprocessing method,data augmentation model,and the LSTM-AE-AM critical scenario recognition model.The DT-based Cellular Vehicle-to-Everything(C-V2X)test system developed based on the research in this thesis and previous work is introduced.