Research And Implementation Of Spatio-Temporally Fine-Grained Traffic Accident Prediction System System

In recent years,with the development of social economy and the explosive growth of the number of automobiles,road safety has become increasingly complex,which leads to an alarming frequency of traffic accidents,seriously restricting the healthy development of the city and profoundly affects the travel safety of citizens.The previous regional and hourly traffic accident prediction has been difficult to meet people’s demand for safe travel.In order to predict traffic accidents at the road level and minute level(i.e.spatio-temporal fine-grained),so as to meet people’s daily travel needs and better assist the traffic management department in designing more reasonable traffic planning.This paper proposes a multi-attention Dynamic Graph Convolution Network model based on Cost Sensitivity Learning to solve the three challenges of this task:(1)how to measure the impact of different factors on traffic accidents.This model uses the Attention Mechanism to score each factor,and uses the score to quantify the degree of influence;(2)How to fit the dynamic spatiotemporal correlation.In this model,the Dynamic Graph Convolution Network is designed to model the dynamic spatial correlation,and the self-attention Mechanism is used to model the dynamic temporal correlation;(3)Zero expansion problem(that is,there are a large number of zero value samples under the granularity of road level and minute level,so that the prediction results of the model are zero).This model introduces the cost sensitivity learning mechanism to increase the cost of false classification of positive samples,so as to accurately mine sparse positive samples.Based on the network model,this paper constructs a web-based prediction system,which realizes three functions:traffic accident early warning,traffic accident heat map display and safe travel path query.Finally,the proposed model is tested on PEMs-Bay and taxiBJ open source data sets.The experimental results show that the prediction accuracy of the network model is better than other baselines in the task.Compared with the current best baseline,F1-score evaluation index in the two data sets is improved by 11.61%and 9.15%respectively.This paper tests the system through test cases,and the test results show that the functional effect of the system meets the expectation.