Recognition and evaluation of driving behavior based on MEMS sensors and machine learning

The aberrant driving behavior can induce vehicle crashes. In order to recognize and evaluate driving behavior for improving the driving performance, real-vehicle experiments were conducted in Banan District of Chongqing, China. First, the built-in Micro-Electro-Mechanical Systems (MEMS) sensor is called through Acceleration, a smartphone application, to collect the data of vehicle driving status. Euler rotation matrix, Fourier transform and low-pass Butterworth filter are used to preprocess collected data. Then, the characteristic variables of driving behavior are selected and their thresholds are calculated to check and filter the samples. A total of 250 samples were selected. Besides, the index set of characteristic variables is established and its dimension is reduced by correlation analysis. Finally, cubic Support Vector Machine (cubic SVM), fine k-Nearest Neighbor (fine kNN) and Gaussian Naïve Bayes (Gaussian NB) classifiers are used to predict and identify driving behavior, and an evaluation mechanism is established based on ride comfort to reflect the aberrant degree of driving behavior. The results show that all 3 classifiers can recognize the sudden deceleration (SD) well and min F1 equal to 93%. But, as for sharp turning (ST), the effect is slightly worse. Cubic SVM has the best training and recognition effect and the overall precision and F1 are higher than fine kNN and Gaussian NB. Sudden deceleration while sharp turning (SDST) has the lowest score and the most aberrant degree. The findings of this study provide a method basis for analyzing and improving driving behaviors.