Risk prediction for cut-ins using multi-driver simulation data and machine learning algorithms: A comparison among decision tree, GBDT and LSTM

The cut-ins (one kind of lane-changing behaviors) have result in severe safety issues, especially at the entrances and exits of urban expressways. Risk prediction and characteristics analysis of cut-ins are part of the essential research for advanced in-vehicle technologies which can reduce crash occurrences. This paper makes some efforts on these purposes. In this paper, twenty-four participants were recruited to conduct the experiments of multi-driver simulation for risky driving data collection. The surrogate measures, Time Exposure Time-to-Collision (TET) and Time Integrated Time-to-collision (TIT) were employed to quantify the risk of cut-ins, then k-means clustering was applied for risk classification of 3 levels. Multiple candidate variables of two kinds were extracted including 10 behavioral variables and 7 driver trait variables. Based on these variables, three prediction models including decision tree (DT), gradient boosting decision tree (GBDT) and long short-term memory (LSTM) are used for predicting the risks of cut-ins. Results from data validity verification show that the data collected from multi-driver simulation experiments is valid compared with real-world data. From results of risk prediction models, the LSTM, with an overall accuracy of 87%, outperforms the GBDT (80.67%) and DT (76.9%). Despite this, this paper also concludes the merits of the DT over the GBDT and LSTM in variable explanation and the results of DT suggest that controlling the proper lane-changing gap and short duration of cut-ins can help reduce risks of cut-ins.