Research On Algorithm Of Vehicle Lane-Changing Trajectory Prediction Based On Social-Convolutional-GAN

In recent years,the rapid growth of traffic volume has caused serious traffic congestion and traffic safety problems,and vehicle lane-changing behavior as a basic micro-driving behavior has greatly affected the road capacity and the safety of vehicle driving.The rationality of the lane change control measures adopted by traffic control is inseparable from the research results of vehicle lane change behavior.Therefore,the study of vehicle lane changing trajectory has always been the focus of theoretical research on traffic flow.However,the traditional lane change trajectory prediction model based on the mathematical model is difficult to accurately describe the real lane change process,and it cannot reflect the influence of the relevant vehicles when changing lanes.Based on the above analysis,this paper takes the freeway vehicle trajectory prediction model as the research object,combined with machine learning technology,a trajectory prediction model based on data driven and fully considering the interaction between the host vehicle and surrounding vehicles when changing lanes is proposed,aiming to accurately describe the actual lane changing process.First of all,this paper combs the domestic and foreign research results of vehicle lane change decision and lane change trajectory prediction,summarizes the three problems existing in the existing data-driven model and proposes the research goals of this article based on this;Many factors affecting the vehicle lane changing behavior,and by comparing the pros and cons of several data-driven trajectory prediction models and applicable scenarios,it is determined to select the LSTM network as the basic structure for processing vehicle lane trajectory data,and then elaborated the model design ideas of this article.Secondly,this paper conducted error analysis and data cleaning on the NGSIM data set in the "Next Generation SIMulation(NGSIM)" project supported by the US Federal Highway Administration,and corrected the longitudinal position,lateral position and vehicle position of the vehicle in stages.The speed information provides reliable experimental data for the training and testing of the model in this paper.Next,this paper proposes a vehicle-change trajectory prediction model based on Social-Convolutional-GAN(SCG)for the limitations of existing models.the model uses the basic framework of generating an adversarial network,and through the game between the generator and the discriminator,it ensures that the generated lane change trajectory is closer to the real data and can better simulate the driving behavior of people;In the design of the generator and discriminator,the LSTM network is used as the encoder and decoder of the lane change trajectory of the vehicle,and the interaction behavior between the vehicles is captured by constructing a convolutional social pool layer in the trajectory generator,solved the problem that the interaction between vehicles is difficult to quantitatively describe during the lane change process,and the introduction of a multi-modal mechanism makes the prediction results more in line with the real lane change scenario;In addition,the strategy gradient descent method is adopted to solve the problem that the gradient of NGSIM discrete data causes the gradient to not propagate in the network.Finally,this paper verifies the validity and robustness of the model by designing relevant experiments.Four existing models were selected as the comparison model in this paper.The experimental results show that the model in this paper has smaller root mean square error and negative log loss function value on the test set;Comparison of the prediction effect of the lane change trajectory between the model in this paper and the single-modal model after removing the lane change intention division module.The results show that the prediction result of the multi-modal model is more in line with the actual lane change scenario;By constructing a new data set,it is verified that the social convolution pool layer can better capture the interaction behavior of the vehicle during the movement;Combined with actual application scenarios,the real-time prediction effects of this model and the two trajectory prediction models V-LSTM and S-LSTM are compared.The results show that the real-time prediction effects of this model are more accurate.The above research results can make up for several major defects of the existing lane change trajectory prediction model,in the end,it can not only provide a more scientific basis for traffic planning and traffic control,but also have far-reaching theoretical and practical significance for the development of microscopic traffic simulation technology and assisted driving technology.