Multisource Accident Datasets-Driven Deep Learning-Based Traffic Accident Portrait for Accident Reasoning

Abstract

Traffic accident data-based portrait plays a vital role in accident cause investigation, relationship reasoning, prevention, and control. The traffic accident data tend to be multisourced with increasingly hidden and complicated accident relationships. The existing reported research focus more on traffic drivers’ measurement of penalty, the relationship among drivers, cars, and dates, etc. How to use multisource data based on deep learning, especially based on the Chinese recent unstructured data and structured data to establish accident portrait for individual and groups of accident drivers, still lacks. Moreover, how to perform multisource accident data label extraction, identity, and relationship extraction are still challenging problems. This paper proposes a multisource accident datasets-driven deep learning-based traffic accident portrait method. Our multisource accident datasets-driven deep learning model is composed of the following three submodels: (1) the structured data accident model using our accident feature-driven bidirectional long short-term memory (Bi-LSTM) and accident feature-driven bidirectional conditional random field (Bi-CRF) model to extract labels, (2) the unstructured traffic accident data model using our accident feature-driven piecewise convolutional neural network (PCNN) model to identify the extracted labels, and (3) the semistructured traffic accident data processing model. Moreover, to solve the problem of how to construct hidden relationship among the multisource accident data, a multisource accident data visualization method based on traffic accident knowledge graph where the accident relational inference algorithm is to complete the hidden relationship between traffic accident data labels is used and then data are visualized using the traffic accident knowledge graph. This paper uses the NER dataset of the People’s Daily and a manually labeled dataset to test the Bi-LSTM + Bi-CRF model, and it acquires the highest scores of 0.9562 and 0.9779 compared with several other models. This paper uses the DuIE dataset and a manually labeled dataset to test the PCNN model, and it acquires the highest scores of 0.9674 and 0.9108 compared with several other models. Experiments verified our model’s merits than other models in regards to accident label extraction, accident identity identification, and accident relationship extraction.