Research On End-to-End Decision-making Of Intelligent Vehicle Based On Spatio-Temporal Recurrent Neural Network

The smart car decision module calculates the decision value according to the input amount of the system to ensure that the smart vehicle runs safely and stably.The traditional smart car decision method uses the lane line information and vehicle information calculated by the vehicle perception module to calculate the decision value,and the decision making depends largely on the input information.Decomposing the decision process of smart cars into lane line detection,vehicle detection,and decision making based on detected driving areas does not guarantee that the entire system obtains the optimal solution.The end-to-end decision-making method based on the deep neural network directly calculates the amount of decision-making based on the input image,and unifies the perceptual cognitive process into the decision-making process.This method of simultaneously optimizing all the processes can achieve better performance and simplify the system structure.Based on this,this paper proposes an end-to-end decision method for smart cars based on spatio-temporal recursive neural network,which can accurately predict the value of the steering wheel angle.The main research work is as follows:In order to reuse network features,strengthen the flow of network information,and further improve the predictability of decision networks.This paper explores the influence of the four feature fusion methods: temporal-spatial feature addition,time-space feature subtraction,spatio-temporal feature multiplication,and spatio-temporal feature cascade to the decision network.Based on the space-time constraint network,this paper proposes an end-to-end decision method for smart cars based on spatio-temporal features fusion recurrent neural network.Next,the principle of the four feature fusion methods is illustrated.Finally,the validity of the spatial-temporal feature fusion network proposed in this paper is verified on the datasets of the Comaai smart car and the Udacity smart car.The test results show that the mean-squared error root of the space-time feature addition method on the Commaai test set is 12.17,further The root mean square error is reduced.The result of the space-time feature addition method on the Udacity test set is also close to the best level in the Udacity competition,and is better than feature addition,feature multiplication,and feature concatenation methods.Aiming at the decision-making problem of smart cars under the actual highway scenario,this paper designs an intelligent vehicle end-to-end decision-making system for expressway scenarios.The feasibility of this decision-making system was verified under the surrounding expressway scene in Changchun City.From the experimental results,the root mean square error of the decision system on the Changchun data set is only 0.098,while the root mean square error of the convolution network is 0.158.Explain that the steering wheel angle predicted by this decision system is closer to the reference value.The decision-making system also performed better in highway cornering scenarios.