| Road traffic safety is a social issue of great concern.Effective and accurate detection of traffic incidents is of great significance to improving the road traffic safety.The traditional expressway traffic incident detection methods are often based on a single source of traffic information,and have many assumptions and constraints on real applications.The ability of incident detection needs to be further improved.Therefore,this paper studies the traffic incident detection algorithms and models using deep learning by fusing the real navigation data and crowdsourced incident report data obtained from the expressways in Beijing.This paper firstly introduces the collection of mobile terminal navigation data and crowdsourced reporting incident data,then explains and analyzes the characteristics of the real data set used.Due to the inevitable errors of the original data and the needs of subsequent analysis,preprocessing operations such as data cleaning,coordinate conversion,and map matching are performed.Secondly,based on the analysis of the impact of the occurrence of traffic incidents on the expressway traffic flow parameters and the behavior of crowdsourced reporting behavior in time and space,the basis of detection is determined.The trajectory data after map matching is spatiotemporal rasterized,the traffic flow parameter characteristics(average speed and density)in the grid are collected,and the crowdsourced incident are clustered by ST-DBSCAN algorithm.The number of reports in the cluster and the mean value of reliability index are calculated.Both features are the input of the detection model.Based on the principle of Convolutional Neural Network(CNN),this paper proposes a traffic incident detection model that fuses two types of heterogeneous data features.A multi-layer convolution structure is constructed to extract high-dimensional features from the traffic flow parameter matrix,and the input of crowdsourced incident reporting features is merged at the layer level.In this paper,a number of comparative experiments are designed to verify the model in different scenarios.Due to the lack of ground truth incidents in the data source,the time and geographical location of incidents are determined by using the speed contour plot method and setting rules.The experimental results show that the proposed model performs ideal detection effects.The integration of crowdsourced reporting information improves the detection ability of the model.Besides,the model can learn the pattern of events determined by rules.This paper also analyzes the impact of incident-free data samples identified by false alarm reports on the model.Finally,aiming at the problems of unbalanced datasets and few-shot learning which often encountered in the incident detection,a deep Siamese convolutional neural network model is proposed.This is a coupled framework constructed by two sub-networks sharing the same weights,and is used to measure the similarity of two input samples to judge whether they come from the same category(i.e.,both incidents or both non-incidents).The sub-network follows the convolutional model proposed above and fuses traffic flow parameter features and crowdsourced incidents features.After the data samples are paired,a large number of new data samples are reconstructed,which makes the training data of the model is multiplied.The proposed improved model massively advances the overall detection results comprehensively,showing a good application prospect in the task of traffic incident detection. |
