Research On Abnormal Warning And Accident Risk Management Of Elevated Road Traffic Based On Deep Learnin

Elevated roads play a very important role as the backbone of urban traffic corridors,and traffic anomalies and accidents on elevated roads often induce deterioration of road levels of service.For elevated road anomalies and accident risks,previous studies are generally based on statistical analysis of static traffic accident data.However,the performance of its real-time detection of traffic anomalies and traffic accidents is relatively poor,and more data support is often needed for road traffic accident risk assessment.In this paper,we first study the characteristics of traffic anomalies,traffic accidents and traffic states on elevated roads,and then propose a series of deep learning solutions to achieve traffic anomaly warning,real-time traffic accident detection and elevated Road Traffic Accident(RTA)risk assessment on elevated roads.This study establishes deep learning models to perform real-time warning of traffic anomalies and detection of traffic accidents based on the available traffic state sequence data features and traffic accident features,and continues to explore the assessment methods of elevated road accident risk based on them.Therefore,this paper focuses on three aspects of real-time warning of traffic anomalies,real-time detection of traffic accidents and traffic accident risk assessment of elevated roads.(1)Real-time warning of traffic anomalies on elevated roads.In this paper,we first combine the available traffic state time series data,and then mine,classify and integrate the data and analyze the spatial and temporal characteristics of traffic accidents.Then,a novel time-series adversarial generative network(TimeGAN)framework,combining both supervised and unsupervised learning,is employed to account for the temporal correlation on the traffic sequence data.This model is optimized for both supervised learning and adversarial training objectives when learning the endogenous space of time series,ensuring that the model can sample historical data while still capturing temporal features.The TimeGAN-based traffic anomaly detection model combines the paradigm flexibility of unsupervised learning with the strong control feature of supervised learning for the training process.Finally,combining with the traffic flow data of Shanghai Yan’an elevated road,the results compare with other early warning model frameworks,the TimeGAN-based traffic anomaly detection model can better achieve early warning of traffic anomalies.(2)Real-time detection of traffic accidents on elevated roads.Firstly,in order to analyze the traffic element sequence data before and after the traffic accident and the characteristics of traffic flow anomalies,the paper establishes a set of traffic anomaly detection model and early warning process based on Seq2Seq autoencoder.Second,an Attention mechanism is introduced to capture important traffic state features based on the Seq2Seq model.Then,by comparing the reconstruction errors of the original and predicted data,and according to the set threshold value,real-time detection of traffic accidents and classification of accident risk levels are achieved.Finally,the feasibility of this real-time traffic accident detection model is demonstrated by using Shanghai Yan’an elevated data and by confusion matrix evaluation approach,and it is found that the average sensitivity of the model reaches 74%and the false positive rate is less than 22%.(3)Traffic accident risk assessment of elevated roads.First,several autoencoders models are designed for traffic accident risk analysis,and each model can achieve rough data classification based on structural errors for the input data.In each round of data classification,all unlabeled data are firstly used as the input of the autoencoders training set.The proportion of normal data in the training set is relatively significantly high,so the corresponding autoencoders can better capture the hidden features of normal traffic flow.Next,the normal data with high confidence level is obtained and the next round of coarse data classification is performed by discriminant criteria.Then,the random forest idea can obtain the normal data set with higher confidence and the corresponding autoencoder model with stronger construction ability,and this model can be applied to reconstruct and discriminate the traffic sequence data in order to finally realize the discrimination of traffic accident risk.In addition,in order to better quantify the road accident risk,this section gives the consideration of feature enhancement for each of the four stages of accident processing.By comparison,the model considering feature enhancement is able to respond more sensitively to prolonged traffic accidents as well as potential traffic anomalies.Based on this,the anomaly detection method combines the reconstruction model and the accident identification model to provide an alternative method for elevated road traffic accident risk and black spot assessment for traffic management.The method overcome the problem of limited accident data size in RTA assessment by deeply mining traffic anomalies.The research in this paper achieves in-depth analysis of elevated traffic accidents,while providing an important active decision basis for elevated road traffic status anomaly warning,real-time traffic accident detection,and three-level linkage of traffic accident risk assessment.