Research On Intelligent Vehicle Steering Decision Based On Deep Learning

In recent years,with the increase in traffic factors such as people,vehicles,and roads in China,road traffic safety problems have also attracted more and more attention.Intelligent vehicles that can effectively improve road traffic safety problems have become a research hotspot.How the vehicle intelligently makes decisions is the research focus of intelligent vehicle technology.In this paper,starting from the intelligent vehicle steering decision control technology,a method of intelligent vehicle steering decision based on deep learning was proposed.The main research contents of this paper are summarized as follows:Firstly,an intelligent vehicle steering decision method was designed.This method transforms the vehicle steering decision problem into a problem of using known environmental information and vehicle steering state information to predict the future steering decision of the vehicle.By designing an end-to-end vehicle steering decision system,the road image sequence and steering wheel angle sequence are used as the input of the system,and the decision model in the system directly outputs the vehicle steering decision value.Secondly,an intelligent vehicle steering decision algorithm model based on deep learning was proposed,which is an encoder-decoder network architecture.The encoder encodes the input road image sequence and vehicle steering wheel angle sequence.The decoder fuses the encoded feature information and calculates the vehicle steering decision value.Inspired by the visual attention and brain memory of human drivers,a spatial attention module based on Conv LSTM and a temporal attention module based on SE-Net were designed and embedded in the visual encoder to increase the model’s attention to the key spatial information in the road image and the key time-step information in the image sequence,while ignoring the interference of other useless information,so as to improve the performance of the model in the road with time-varying environmental factors.Then,the experimental dataset was made based on the comma2k19 public dataset,and the algorithm model was built using the Tensorflow.The grid search method was used to complete the model’s hyperparameter adjustment.Through comparative experiments with advanced models,the decision-making performance of the model was verified qualitatively and quantitatively.The proposed model was trained using different numbers of training set samples,and the impact of the size of the training set on the performance of the decision-making model was analyzed.Finally,the spatial attention and temporal attention of the model were visualized to further explain the learning ability and decision mechanism of the model,and to show the internal calculation process of the model’s decision-making.