Research On Path Planning And Dispatching Of Unmanned Vehicles Facing The Real Environment

With the development of driverless technology and people’s attention,driverless vehicles perhaps become the mainstream model of online car-hailing platform in the future.Only by using the real-time order demand and the current situation data of the road network is difficult to operate driverless vehicles efficiently.However,the historical trajectory and order data of online car-hailing fully reflected the law of urban road network operation and residents’ travel.Therefore,this study continues the research work of intelligent operation of unmanned vehicles,including traffic light detection,path planning and scheduling,with combining the above data with the actual environment data to.Firstly,for reducing the map cost of unmanned vehicle operation,this study used trajectory data to update the map with low cost to add the missing traffic information in Open Street Map（OSM）.Secondly,this study combined point of interest（POI）information and weather information to predict the road conditions,and calculated the optimal path to shorten the driving time of unmanned vehicles.Finally,this study reduced the operation cost of unmanned vehicles by using the online car order data and related data to predict the demand,and calculates the optimal scheduling strategy of unmanned vehicles to balance the supply and demand,so as to effectively.The following is the main work of this study:1.Detecting traffic light using traffic data（DTL）.By analyzing the track data,the model extracts the vehicle track speed sequence,the type of intersection,the traffic flow and the number of roads connected to the intersection to detect the traffic lights.This study first obtains the intersection position and traffic light mark of some areas according to the road network map,and then makes incremental matching according to the intersection position and vehicle speed to determine the influence range of traffic light,so as to further mark the category of traffic light;Then,the training set is segmented,and a multi-input network built with density module and LSTM module（DLSTM）is designed according to the characteristics of the data;Finally,all the track data are used to detect the traffic lights of the whole city.The experimental results show that the AUC value under ROC curve is 0.95,and the performance of this method is the best.In addition,the traffic light information can also improve the accuracy of the travel time estimation in the path planning and the arrival time estimation in the scheduling.2.Learning path planning using deep Q network（LPQ）based on road condition prediction.Firstly,based on the idea of prediction and planning,this study designed the framework of unmanned vehicle driving system for actual environment;Then,a fast global path planning model LPQ is designed based on deep Q-learning and deep prediction network.In order to improve the driving efficiency of unmanned vehicles,the model predicts the future road travel time through the data of road space-time,weather and poi distribution,and uses DQN algorithm to solve the shortest travel time path.Finally,this paper evaluates and verifies the ability of the model on the public data set and the actual urban road network map.Experimental results show that compared with other common methods,the travel time of this method is reduced by 17.97%.3.Learning dispatch policy using Q-table（LDQ）based on demand forecasting.The model predicts the future demand of each region through the historical order data,weather,time and poi distribution of online car-hailing,and dispatches the unmanned car to the designated area in advance to deal with the upcoming orders,so as to reduce the no-load passenger search time of the unmanned car.Firstly,the city map is divided into several regions according to the order time window;Then,according to a certain time interval,the number of orders in a data set is counted as the demand data;Then,according to the characteristics of the data,this study designed the DCLSTM（multi-input network build with Density module,CNN module and LSTM module）network,trained the network to get the regression predictor,which is used to predict the order of the carrying area;Finally,this study used Q-learning algorithm to calculate the best scheduling scheme according to the predicted bus demand.Experiments show that compared with the common methods,the cumulative revenue of the model LDQ can be increased by up to 5%.