Data Driven Mixed Traffic Simulation

As an important technical analysis method for modern urban road traffic and one of the main forms of test environments for autonomous vehicles,mixed traffic simulation has great theoretical significance and application values.The force-based traffic simulation is a representative method of mixed traffic simulation,and it can accurately simulate the complex behaviors of various types of traffic agents through a simple and unified generalization model.However,due to the complexity and strong correlation of model parameters,a good simulation result mostly requires frequent parameter adjustment,which greatly increases labor costsTo solve the problem mentioned above,this paper proposes a new data-driven method for social force-based mixed traffic model to realize parameter estimation.The method constructs the error equation with the coordinates of the specified object in real and simulated mixed traffic traj ectory,and maps the error value into the fitness of genetic algorithm,then obtains the model parameters of the object through iterative optimization procedure.In order to avoid systematic overestimation of absolute error measure for large errors where the traffic object moves at a high speed as well as the inconsistent sensitivity of the relative error measure for trajectories of different length,a new objective function is proposed in this paper which combines the characteristics of the two measures.Meanwhile,to reduce the possibility of converging to a local optimum,this paper utilizes adaptive genetic algorithm,where the crossover and mutation probability of each individual will be modified by quantifying the fitting of temporal parameter population's fitness,and the elitist strategy is introduced to improve the convergence speed by retaining the optimal solution set of each generation in iteration procedure.In addition,this paper proposes a series of model optimization schemes for existing models referring to the behavior of agents in real traffic scenarios,including the modification of road constraints,consideration of individual sensitivity to different forces,introduction of perspective influence factors,and so on.Based on the optimized model,road information extraction,spatial transformation and calculation of motion information such as target unit displacement,velocity and acceleration,are performed sequentially on real traffic dataIn order to prove the applicability and feasibility of the model,the paper simulates a mixed traffic scenario which contains lane changing of vehicles,road crossing of pedestrian,avoidance of non-motor vehicles and so on.At the same time,the parameter values obtained by optimization are employed in further simulation,where the trajectory coordinates,velocity-time distribution are compared with real data to verify the usability of the solution,error-iterations and error distribution of the whole scene are statistically analyzed for method performance.