Research On Behavior Decision-making Method Of Autonomous Vehicles Based On Improved GRU And SVM

With the rapid development of artificial intelligence and big data,autonomous driving technology has set off a research boom in the world in recent years.Autonomous driving technology is mainly divided into three areas: environment perception,behavior decision-making,and path planning.Behavior decision-making algorithms generate driving commands based on road condition information to ensure that autonomous vehicles behave like vehicles driven by experienced drivers.Therefore,it is necessary to develop behavior decision-making algorithms with high decision accuracy and high real-time performance.Existing behavior decision-making methods are divided into mathematical modeling methods and machine learning methods.However,disadvantages are existed in both of two types of methods.The mathematical modeling methods establish state transition models based on logic rules and build vehicle behavior models by calculating risk indicators.The establishment of models is generally based on complex mathematical formulas.Therefore,mathematical modeling decision-making algorithms are usually computationally expensive.In addition,mathematical modeling methods are limited to specific scenarios,leading to weak generalization ability.The machine learning methods improve the generalization ability of models through extensive training.However,the machine learning methods only generate decisions based on the input at the current moment without considering the historical state of vehicle,resulting in low time correlation.Therefore,the pursuit of high decision-making accuracy,strong time correlation,and strong generalization ability is a hot and difficult point of research scholars for many years.In this paper,a systematic research on the behavior decision-making method of autonomous vehicles is discussed,and a behavior decision-making method of autonomous vehicles based on improved GRU and SVM is proposed.The method includes the following three innovations:(1)Aiming at the problem of weak generalization ability of mathematical modeling methods and the problem of weak time correlation of machine learning methods,a feature extraction method based on deep learning methods is proposed.According to the different lane and vehicle position,an improved GRU(Gated Recurrent Unit)is designed to extract the features of the self-vehicle and surrounding vehicles,and the two feature extraction results are merged in each iteration,which improves the applicability of the decision results.(2)Aiming at the problem of low classification accuracy of traditional classification algorithms,a decision generation method based on Support Vector Machine(SVM)and Moth-Flame Optimization Algorithm(MFO)is proposed.Support Vector Machine is a Machine Learning algorithm widely used to solve high-latitude multi-classification problems.Moth-Flame Optimization Algorithm is a meta-heuristic optimization algorithm used to optimize parameters.In this way,the algorithm in this paper improves the accuracy of decision-making results.(3)Aiming at the problem that MFO is easy to fall into the a local optimum solution and converges slowly,an enhanced MFO is proposed based on the Levy flight and an adaptive weight,improving the ability of global searching and convergence speed.This paper conducts experiments on the proposed algorithm on the NGSIM(Next Generation Simulation)data set.The proposed algorithm is compared with several latest behavior decision-making algorithms.The result shows that the algorithm in this paper performs better on accuracy rate of decision results.Behavior decision-making algorithm of autonomous vehicles is the research focus of this article.In this paper,the idea of module division is applied.First,for feature extraction module,an improved GRU is used to extract vehicle driving features.Secondly,Levy flight and an adaptive weight are introduced to MFO to improve the optimization ability.Finally,for decision generation module,an enhanced MFO is combined to SVM with strong classification performance to generate decision-making results.