Research On Intelligent Driving Decision Based On The Cognition Of Traffic Situation

With the continuous increase of car ownership,traffic safety and traffic con-gestion are becoming more and more serious.In order to solve these problems,it depends on the new generation of Intelligent Transportation Systems(ITS).With the development of ITS,autonomous driving is becoming the develop-ment direction of automobiles.Due to the cognitive limitations of current au-tonomous vehicles,which are affected by perceived accuracy,distance and cost,etc,it is impossible to solve the driving problems of safety and efficiency only relying on individual intelligence.Therefore,it is necessary to use vehicle-road collaboration technology to transform from individual intelligence to swarm intelligence.Through swarm intelligence collaboration,it solves "over-the-horizon cognition" and "knowledge sharing" problems of individual intelli-gence,and realizes the functions containing complex environmental perception,intelligent decision-making,collaborative control and execution,so as to achieve a safe,efficient,energy-saving,and comfortable "Intelligent driving".Further intellectualization of driving-oriented behavior decision making is re-alized.Based on the development direction of automobile intelligence and ve-hicle-road synergy,this paper focuses on the study of traffic situation recogni-tion,individual decision-making and group decision-making.The main inno-vations are as follows:1)In view of the complex and changeable traffic environment,a fine-grained traffic situation cognition method based GC-LSTM for road sections is proposed,based on that the traffic situation of coarse-grained cognition cannot support accurate routing decision-making problems.Specifically,the road to-pology network is abstracted into a graph network at first.Considering the im-pact of different road attributes and different intersection types on traffic flow,it is proposed to integrate the road attributes into the graph signal,and design different convolution kernels for intersections with different in-degrees.Graph convolutional networks are used to learn the spatial correlation of traffic flow.Then,a long and short-term memory network is utilized to learn further to cap-ture the time dependence of traffic flow,based on the capture of spatiotemporal relationship of traffic flow,the change trend of traffic state is recognized.Ex-perimental verification based on the real traffic data sets shows that GC-LSTM can effectively reduce the average prediction error about MAE in traffic situa-tion prediction.Compared with FC-LSTM,ConvLSTM and STSGCN,the av-erage prediction error about MAE of GC-LSTM is reduced by 30.6%,28.1%and 12.37%,respectively.2)Aiming at the problem of route planning based on historical laws and current road traffic conditions that does not conform to "future expectations",causing deviations in route planning and low user compliance,a road topology-oriented autonomous navigation decision-making method based on 2r-GVIN is proposed.Specifically,the traffic prediction network is used to recognize the traffic situation of each road section in the whole area,and to provide "predic-tion" rewards for the decision-making network.Then,build a Markov decision-making process oriented to driving planning,combined with the "current" traf-fic states rewards,a double reward generalized value iterative decision network is designed,and a "color value map" is generated for navigation planning to support efficient autonomous driving decisions.Experimental verification based on real traffic data sets shows that route planning based on 2r-GVIN can effectively reduce the actual travel time.Compared with LSTM-VIN and R-VIN,the travel time of 2r-GVIN is reduced by 10.4%and 5.46%respectively.Meanwhile,the accuracy rate and success rate of route planning are increased by 21.5%and 17.57%compared with LSTM-VIN,and 5.82%and 10.45%compared with R-VIN.3)Aiming at the problem that driving decision-making in a dynamic traffic environment with uncertainty and linkage is difficult to take into account the global and local optimal,resulting in low traffic efficiency,a cross-domain col-laborative decision-making method based on vehicle-road collaboration,2L-CoV,is proposed.Specifically,in order to enable collaborative interaction be-tween vehicles,a distributed collaborative framework supporting large-scale virtual vehicle interactive computing is designed.Considering the characteris-tics of the global traffic state and the adjacency relationship between regions,then an improved back pressure algorithm is proposed to achieve rapid guid-ance for traffic flows.Also,considering the influence of driving behavior pref-erences on decision-making,a reinforcement learning method based on domi-nant strategy for game evolution is proposed to complete local vehicle routing planning.The simulation results show that compared with SP,SL,BPSP and BPSL,the global cooperative scheduling based on 2L-CoV can effectively im-prove the traffic efficiency,and the average throughput of the network can be increased by 56.88%,25.65%,32.82%and 7.93%,respectively.The average travel time decreased by 25.6%,12.36%,14.4%and 7.6%,respectively.