A Hybrid Method for Traffic Incident Detection Using Random Forest-Recursive Feature Elimination and Long Short-Term Memory Network With Bayesian Optimization Algorithm

• Published in: IEEE access Vol. 9; pp. 1219 - 1232
• Main Authors: Shang, Qiang, Feng, Linlin, Gao, Song
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Bayesian analysis; Bayesian optimization algorithm; Detectors; Feature extraction; feature selection; Hybrid systems; Intelligent transportation systems; long short-term memory network; Machine learning algorithms; Optimization; Optimization algorithms; Prediction algorithms; Roads; Sampling methods; Short term; Support vector machines; Traffic incident detection; Transportation networks; Volume measurement
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3047340

Automatic Incident Detection (AID) is an important part of Intelligent Transportation Systems (ITS). A hybrid AID method using Random Forest-Recursive Feature Elimination (RF-RFE) algorithm and Long-Short Term Memory (LSTM) network optimized by Bayesian Optimization Algorithm (BOA) is proposed in this article. Firstly, a relatively comprehensive set of initial variables is constructed using basic traffic variables and their combinations. Secondly, feature variables are selected from the initial variables using the RF-RFE algorithm. Then, the feature variables are used for training the LSTM network, and the hyper-parameters of the LSTM network are optimized by BOA. In addition, Synthetic Minority Over-Sampling Technique (SMOTE) is employed to solve the problem of imbalance between incident sample size and non-incident sample size. Finally, experiments are conducted using real-world data to test performance of the proposed method and compare with several state-of-the-art AID methods on multiple evaluation criteria. The experimental results illustrate that the proposed method achieved superior performance with respect to almost all the evaluation criteria. It also shows that the proposed method is promising for dealing with the problems of imbalance and small sample size of traffic incident data.