Research On Safety Situation Identification In Car-Following

This paper first analyzes the main influencing factors of vehicle braking process and braking distance,uses Carsim vehicle dynamics simulation software to simulate vehicle braking distance under different road conditions and vehicle braking initial speed,and uses improved particle swarm optimization to optimize the limit learning machine The algorithm builds a vehicle braking distance prediction model.Based on the braking distance between the front and rear vehicles and the desired safety distance after parking,the calculation method of the following distance of the vehicle is obtained.Secondly,using NGSIM vehicle trajectory data to train and build two vehicle-following models based on deep learning.The model can simulate the driver’s memory effect and the model is completely data-driven,minimizing human interference.Improve the network structure,activation function,and hyper-parameters of the model to solve the problems of overfitting and disappearance of gradients.Observation results show that using deep learning networks and time series data can better extract the deep features of the data.When using this data set,the best prediction results can be obtained by inputting 9 ～ 10 s historical time length data.In order to identify the vehicle’s car-following safety situation,a deep learning car-following model is used to predict the driving speed of the car at a future time point,and this speed is used as the initial braking speed of the vehicle to predict the vehicle braking distance.The safety distance that the car should maintain.Use this distance to calibrate the safety status of the NGSIM vehicle trajectory data set.Then,the transformed deep learning car-following model is subjected to migration training using the calibrated data set,so that the model can classify and identify the vehicle’s safety situation in real time.Based on the sample imbalance problem,the sample weights of different categories in the model loss function are improved.Finally,the test data is used to simulate the real prediction accuracy of the model using a special evaluation method for the imbalance of data samples.The ROC curve is drawn to prove that the improved temporal convolutional network model can be guaranteed under the setting of multiple classification thresholds.Higher classification accuracy.Based on the above research results,the improved temporal convolution network identification model proposed in this paper has a better performance in vehicle safety situation identification based on measured traffic data,and can be used as an effective identification scheme for forward collision avoidance systems of automobiles.