Research And Verification Of Testing Scenario Generation Methods For Pedestrian Path Conflict In Autonomous Driving Cars

Conducting scientifically sound testing and verification is a prerequisite for the largescale application of autonomous driving cars.Virtual simulation has the advantages of efficiency,safety,and low-cost.And it is an important method for testing autonomous driving cars.Pedestrian intrusion into the lanes and path collision with vehicles is a common and typical dangerous scenario in a straight-road environment.It is of great significance to verify the safety of autonomous vehicles to test autonomous vehicles in this scenario.The thesis relies on the national key research and development plan project "In the Loop Testing Technology for Autonomous Vehicle Traffic Based on Digital Twins"(2021YFB2501204).It focuses on the pedestrian intrusion into the lane scenario in a straight-road environment within a school.Through video capture,pedestrian detection and trajectory tracking,real pedestrian trajectories in this scenario were extracted and a pedestrian trajectory dataset in this scenario was built.Based on the generative adversarial network,a pedestrian trajectory generation model was constructed to achieve automatic generation of pedestrian trajectories and increase the diversity of trajectory data in the scenario set.Based on this,a pedestrian intrusion into the lane test scenario was constructed and applied to virtual simulation testing of autonomous driving cars to verify the performance of different automatic emergency braking algorithms in the autonomous driving system under this scenario.The thesis’ s research content includes four aspects:(1)Based on collected real-scenario video data,pedestrian detection and trajectory tracking models were constructed.Firstly,the thesis built a YOLOv5 object detection network based on the characteristics of real-scenario data to recognize dynamic pedestrians in realscenario videos.Secondly,pedestrian trajectory tracking was implemented using the Deep SORT pedestrian tracking algorithm.By extracting the motion trajectories of pedestrians intruding into the lane in the real-scenario data,the trajectory dataset was established.The YOLOv5 and Deep SORT methods used in the thesis have the advantages of high accuracy and good robustness,which can accurately extract pedestrian trajectory data.(2)A pedestrian trajectory generation model based on Time-series GAN was constructed.Using the real pedestrian trajectory dataset as the training dataset for Time-series GAN,a pedestrian trajectory generation model for the pedestrian intruding into the lane scenario was established to generate pedestrian trajectories.(3)Combining the static scenario elements in real-world scenarios,the thesis constructed a virtual testing scenario for the pedestrian intruding into the lane scenario based on Pre Scan and Matlab/Simulink.The thesis designed a tool to import real trajectories into Pre Scan,which enabled the replication of real and generated trajectories in the testing scenario.Referring to the European New Car Assessment Program and China New Car Assessment Program,the thesis designed three typical test scenarios for pedestrian crossing.(4)The thesis conducted virtual simulation testing under three typical test scenarios and compared the effects of different automatic emergency braking algorithms on the safety and effectiveness of autonomous vehicles.Using Pre Scan,the thesis set up a pedestrian intruding into the lane test scenario and used Matlab/Simulink to construct three automatic emergency braking algorithms: TTC,Mazda,and Honda.The thesis validated and compared the performance of three algorithms under three typical test scenarios.The simulation test results showed that the safety and effectiveness of the TTC model were the best in the pedestrian intruding into the lane scenario.The thesis used video data to construct models for pedestrian trajectory detection,tracking,and crossing lane trajectory generation,and obtained a dataset for pedestrian crossing scenarios.The import of real pedestrian trajectories into virtual simulation testing scenarios was realized and the application effects of different automatic emergency braking algorithms in this scenario was validated.The research results have important significance for conducting tests on path conflicts between autonomous vehicles and pedestrians.