Sensor-Based Risk Perception for Drivers Under Adverse Environment

Abstract

Due to factors such as snow and ice impeding drivers’ vision, automobile accidents significantly rise during winter months. This study sets forth an automated evaluation network of the Risk Perceived Ability (RPA) for motorists driving on the freeway in snow and ice environments, using a deep learning approach and the Rough Sets technique. First, a natural driving experiment involving thirteen licensed drivers was conducted on a freeway in Jilin, China, with a dangerous point set prior to the start of the experiment. Then multi-sensor (eye-trackers, mini-cameras, and speed detectors) apparatuses, collecting both images and numerical data, were utilized. Afterward, Restricted Boltzmann Machine (RBM) was used to develop a deep belief network (DBN) along with training procedures. Rough Sets technique was added as judgment in output layer of the DBN. Finally, fixation duration, pupil size, changes in speed, etc., were used as input impact factors and the perception conditions were used as output variables to train the network. Furthermore, after comparing the DBN-based risk perception ability network with Naïve Bayes and BP-ANN (artificial neural networks with back propagations), it showed that the results indicate that the developed DBN not only outperforms both Naïve Bayes and BP-ANN, but also improves the accuracy of perceiving risky conditions to 90% - 95% in HSRs. This approach can provide reference for the design of hazard detection systems of partially-automatic vehicles.