Research On Safety Evaluating Methods For Driverless Trams Based On LSTM

As an important means of urban public transportation,promotion and popularization of driverless modern trams have become consensus of the governments and scholars at home and abroad to solve urban traffic problems.Safety is the core content of driverless technology and the prerequisite for application and promotion of the driverless trams.At present,it is impossible to apply automobile data directly for the trams research,since volume,weight,formation,braking/deceleration,and other parameters of trams are quite different from those of automobiles.How to apply deep learning method to the driverless trams’ safety evaluation remains to be studied and verified.Existing safety evaluative methods based on traj ectory prediction have poor adaptability to weather,road conditions and other environmental factors,and are difficult to reflect correlations between vehicles such as vehicle following and overtaking.For the current situation that no tram’s dataset can be found in open source datasets at home and abroad,Simulation datasets of the trams in traffic scenes are built up in this paper according to characteristics of the trams,body and driving characteristics.Deep learning and neural network are applied to trajectory prediction and safety evaluation of the trams.In order to solve the shortcomings of existing trajectory predictive-based safety evaluation methods,new tram safety evaluative methods based on interactive analysis are explored by combining deep neural network technology and kinematics formula.Main research works in this paper as follows:1.Road simulative environments are established with real road information as the topology,based on characteristics in vehicle topography and driving of the trams.The traffic scenes are enriched by adding vehicles and pedestrians as target obj ects,and the simulative datasets of tram traffic scene are built by setting driving speed zones and random acceleration intervals to increase the randomness of simulative data.2.According to time sequence characteristics of motion trajectory,the deep learning and neural network are applied to the safety evaluation of the trams,and the trajectory predictive methods based on long-term short-term Memory network(LSTM)are verified.Influences of learning parameters and model structures of LSTM-model on accuracy of prediction and convergence rate of training process are studied and analyzed.3.Collision detection method based on point domain is adopted to deal with influences of environmental factors and correlation between vehicles on safety of trams in actual traff-ic scenes.By combining LSTM deep neural network with vehicle safety distance formula,a tram safety evaluative method based on interactive analysis is explored,and the feasibility of the method is verified by experiments.Experiments show that the LSTM-based interactive safety evaluative model has better prediction accuracy and model convergence speed,which verifies the feasibility of the method.The influences of different parameters on model performance are obtained and the optimal model is found.