Research On The Vehicle Trajectory Generation Method Based On Variational Auto-encoder In Typical Traffic Scenarios

Autonomous vehicles need to go through a large number of simulation tests to ensure the safety of each functional module of the vehicles before they are officially launched on the market.To build test scenarios that reflect the real traffic dynamics,the simulation testing requires vehicle trajectory datasets to provide support.However,the acquisition of vehicle trajectory data usually requires a lot of manpower and material resources.Therefore,there is an urgent need to conduct research on vehicle trajectory generation methods in typical scenarios to save the cost of data collection and provide more available data for autonomous driving simulation testing.In recent years,the development of variational auto-encoders(VAE)has brought opportunities for trajectory generation research.VAE learns feature distribution from natural driving trajectories,generating diverse trajectories and providing more data for autonomous driving simulation testing.However,when the VAE model is applied to trajectory generation,there are still some challenges:(1)how to integrate the vehicle kinematics mechanism into the data-driven generative model to randomly generate vehicle trajectories that satisfy the kinematic constraints;(2)how to design trajectory targeting generation models that consider the interaction between two vehicles under given vehicle driving state conditions;(3)the Safety-critical trajectory(Safety-critical trajectories)are scarce in the dataset.How to generate low-probability safety-critical trajectories that are challenging for autonomous driving test vehicles while ensuring the authenticity of trajectories is an important problem that needs to be solved.In response to the above challenges,this paper focuses on typical traffic scenarios such as following and lane changing,and conducts research on vehicle trajectory generation methods.The main research contents are as follows:(1)Aiming at the random generation of trajectories satisfying kinematic constraints,a trajectory generation method based on a data-mechanism mixture is proposed.This method uses the variational autoencoder model to capture the natural driving characteristics of the trajectory in the data and establishes a two-degree-offreedom vehicle model to describe the motion mechanism of the trajectory.By incorporating the motion mechanism into the data-driven generation model through the addition of kinematic constraints during the trajectory decoding process,the randomly generated trajectories satisfy reasonable kinematic relationships.(2)Aiming at the problem of trajectory generation from random to targeted,a trajectory generation method considering vehicle interaction is proposed.This method uses the trajectory driving state as a condition and constructs a target-oriented trajectory generation model based on the conditional variational autoencoders(CVAE).It also considers vehicle interaction relationships,designs a parallel framework,and establishes a Dual-CVAE model to generate trajectories that satisfy the interaction relationship of two vehicles under given driving conditions.(3)Aiming at the generation of safety-critical trajectories with low probability,a safety-critical trajectory generation method considering traffic prior distribution is proposed.The trajectory generation problem is described as an optimization problem,and an adversarial optimization function is designed to stimulate the generation of trajectories with small probabilities in the tail distribution.The trajectories are then constrained based on the traffic prior distributions to achieve adversarial and realistic safe and critical trajectory generation.The effectiveness of the three types of trajectory generation methods has been validated in typical scenarios,and they can generate normal driving trajectories with natural driving characteristics and safe and critical trajectories with low probabilities.These methods can provide data support for automated driving scenario testing.