Vehicle Collision Event Prediction Based On Multi-view Collaborative Training

Car crash risk prediction is of great importance in traffic safety management.According to statistics,millions of people die each year from road traffic accidents around the world.Bad driving behaviors are the main cause of traffic accidents.Car crash risk prediction is a very realistic research.On the one hand,it can serve the construction of smart cities.By predicting the risk of car driving,drivers can be reminded of driving safety,which can reduce road accidents and it can help urban road management.On the other hand,it can also provide the driver's driving safety assessment results for auto insurance service providers,can help auto insurance service providers solve the problem of "adverse selection",and help auto insurance service providers to screen out high-quality driver customers.The concessionary price of auto insurance achieves a win-win situation.With the widespread application of Internet of vehicles(Iov)and the increasingly mature big data technology of Iov,scholars have begun to use the Iov data to quantify the driving behavior of drivers and evaluate and predict vehicle driving risks.However,the previous research extracts features related to car driving speed with limited data,and has not fully explored other data rich in Iov.The data set used in this article includes 326 crashed cars,2353 safe cars,and a total of 2779 cars.Through empirical analysis,it is found that the car crash event is not only related to the speed features studied by previous scholars,but also to the spatialtemporal features,such as whether the car is often driven at night or whether the car is driven to some unfamiliar areas.Therefore,this paper not only extracts the speed features related to the car crash event from the car's speed acceleration data,but also extracts the spatial-temporal features related to the car crash event from the car trajectory data.Therefore,this paper intends to use the historical Iov data to evaluate the driving behavior of the driver,so as to predict the risk of a car crash event in the future.The two parts of data used in this paper are one part of the data,one is the latitude and longitude data of the GPS positioning of the car,that is,the trajectory data of the car,and the other is the data of the speed acceleration during the car driving.Furthermore,based on these two pieces of data,their prediction models are respectively established,which are "spatial-temporal features classifier" and "speeds features classifier".Each classifier has three parts of input: artificial features,input maps of all data,and input maps of night data.Artificial features are some features related to car crash events extracted according to previous scholars' research results and daily life experiences.The input map of all data is initialized into the input map with all their respective data.We intend to use deep neural networks to input maps from the input map.Extract some hidden features to predict the problem.Similarly,the input map of night data is an input map constructed by separately separating night data from all data.These two models are trained using the model results of the deep residual network.Through empirical analysis,it is found that these two classifiers have a certain ability to predict car crash events.However,this paper does not use the more traditional method of concatenating features to predict the problem.The co-training method makes use of the complementary prediction relationship of these two parts to make up for their respective prediction capabilities for some difficult-to-predict samples.In addition,co-training can add pseudo labels to unlabeled samples,put the samples with pseudo labels into the training set,increase the number of samples in the training set,and alleviate the situation of model underfitting.Finally,we analyze the experimental results and some parameters in the experiment.Through experimental comparison,we find that multi-view input will enhance the prediction effect of the model.The prediction effect of the model after cotraining is better than the simple fusion of the two features.Our parameter analysis of co-training found that the more pseudo-sample cars added each round,the Top-n Precision results will decrease.This is because co-training will add a part of the wrong samples into the training set,which are the wrong training samples.The impact on Topn Precision is greater.At the same time,we also found that within a certain number of iterations,the more iterations of co-training,the better the AUC results.