Online Driving States Monitoring Using Fusion of Multi-sensor Information

In this paper, an online methodology for the detection of unsafe driving states while driving is presented. The detection is based on the multi-sensor approaches, including gyrometer, accelerometer, radar, video and so on. Various information comes from both the ego vehicle and its surroundings are fused to gain a comprehensive understanding of driving situations. Using subspace modeling techniques, we propose an unsupervised learning algorithm to perform the unsafe states detection. The feature space are decomposed into the normal and anomalous subspace, where the normal space are assumed as the major components of the driving patterns, and significant deviations from the modeled normal subspace are signaled as unsafe states. In addition, the algorithm works in a real-time way incorporating a implementation of sliding window, which enable the method to adapt over time to address changes in the new emerged driving situations. We have implemented our algorithm with a prototype system installed in a transit bus, validations are performed in real driving situations. Our experimental results demonstrate the effectiveness of the approach on forward risk predication. We gain a timely predication while with a low false positive when there occurs conflicts between the ego vehicle and front vehicles.