Detection Of Vehicle Safety-critical Events Based On Multi-Features In Network Environment

Vehicle safety-critical events refer to the detection of abnormal conditions of the vehicle by integrating driving data and its surrounding environment.This technology is extremely important for reducing accident losses and ensuring driver safety.The traditional recognition of safety-critical events is affected by the sensor’s sensing range,and only the local features of the vehicle dynamics data are used to recognize safety-critical events.This method has certain limitations.This paper integrates multi-dimensional features such as vehicle dynamics data and road fluctuation parameters obtained from other vehicles to study the detection of safety-critical events in network environments.This article mainly uses the SUMO simulator to study the multi-dimensional features under the connected and automated vehicle environment and uses machine learning and other related fusion algorithms to establish safety-critical events model.Finally,it is verified with natural driving data.This article is based on the National key research and development program(2017YFC0804800).The main contribution of the dissertation is depicted as follows.1)Feature analysis and extraction of vehicle safety-critical events detection in network environment.First,the SUMO simulator is used to simulate two kinds of road environments of straightline network and intersection road network.According to the different road network environments,traffic accidents are simulated,and the accident dataset is extracted and consider as dataset of the theoretical experiment in this dissertation.Secondly,the coefficient of variation,the quartile coefficient of variation,the percent of extreme values,and other statistical indicators are used to measure the whole road.The correlation analysis shows that there is a certain correlation between road characteristics and accidents.Thirdly,the initial feature set of this paper is composed of vehicle characteristics including speed,acceleration,speed difference,the distance between cars,and road fluctuation characteristics,with a total of 39 dimensions.Finally,two feature selection algorithms,random forest,and GBDT,are used to rank the importance of features under different road networks to form the optimal feature subsets of the two road sections.2)Detection model of safety-critical events in a network environmentFirstly,three feature sets are established,they are single-vehicle feature set,optimal feature subset based on random forest,and optimal feature subset based on GBDT.Three different classification fusion algorithms including logistic regression,support vector machine,and Adaboost are used to train and model the three feature sets,nine early warning models were formed.Then,using the common evaluation indicators of classification algorithms such as accuracy rate and precision rate to evaluate nine models and combined with predicting the probability of collision,the optimal model is selected.Finally,the experiment proves that the model proposed in this paper has better performance than the single-vehicle feature model.Simultaneously,it is proved that the GBDT-SVM model has the best effect based on the straight-line network.The accuracy rate reached 92.5%,the recall rate reached 92.9%,and the F1 Score was 92.3%.It proves that the GBDT-SVM model has the best effect based on the road network of the intersection.The accuracy rate is 91.7%,the precision rate is 85.8%,the recall rate is 96.6%,the F1 Score is 90.9%,and the AUC area is 0.924.3)Validation of multi-feature safety-critical events recognition model based on an open natural driving datasetThe DAS2 data of the United States SPMD project have been used in this experiment.Firstly,the original natural driving dataset is processed to meet the accuracy of the experiment,including data filling,data deduplication,and other operations.Secondly,the threshold method is used to find the possible abnormal fragments and manual validation.And the article also uses Google Earth to restore the location of the vehicle,and divide the original abnormal events into21 straight road abnormal events and 38 intersection abnormal events.Then,a 36-dimensional feature set is established for the dataset.Finally,the optimal feature selection algorithm described above is used to filter them and the classification fusion algorithm is used to detect safety-critical events.Through experiments,it has been proven that the accuracy rate of GBDTSVM model is 96.6%,the accuracy rate is 93.2%,the recall rate is 89.1%,the F1 Score is 91.1%,and the AUC is 0.937.In term of intersection road network data,the accuracy rate of GBDTSVM model is 95.4%,the accuracy rate is 85.7%,the recall rate is 86.5%,the F1 Score is 86.1%,and the AUC area is 0.919.