Multi-CNN and decision tree based driving behavior evaluation

Abstract

Driving behavior is directly related to the lives and property safety of the drivers and others, good driving behavior can not only reduce the accident rate, but also reduce the driving risk. In this paper, an effective driving behavior evaluation method is proposed. Features integration is very important, we propose the multi-CNN architecture, it has a higher prediction accuracy. Convolutional neural network is time-consuming and computation intensive, a dynamic fixed point compression method is applied in our system, smaller model size and faster speed can be achieved while the accuracy is high. The lanes, cars and pedestrians on the road are detected, meanwhile, the distance between the host car and the nearest car in front of it is calculated. These data are predicted by a trained Gradient Boosting Decision Tree model, the prediction result is a driving score that can reflect the driver's driving behavior is good or bad. The root mean square error of our model is 1.9, which has a high accuracy and is useful in practice.