| In this Thesis, we focus on the analysis of driver lane-changing behavior based on the fact that lane changing is a ubiquitous driving maneuver in common driving environments and regarded as the most critical driving intention. Therefore, lane changing as a case study for driving intention recognition is introduced in this study. Our methodology is to employ machine learning method i.e., support vector machine, to the classification of driving intentions using vehicle performance data and driver eye gaze data from the measurement of the driving tasks (i.e., lane following and lane changing) in the well-designed simulation environment. To improve recognition performance, this thesis illustrates the use of entropy based fusion method to discard those largely negative dependent data to decrease the redundancy level of input data. Based on entropy correlation coefficient analysis, heading angle, as one kind of performance data, is harmful for recognition result. Incorporation of eye gaze data enhances the recognition performance. Introduced by three-stage nested design, experiments are executed to screen out the best "supplies" for pattern learning. Final results show that feature set fused by steering angle, gas pressure, velocity and acceleration performance data as well as eye gaze data achieves 88.78% accuracy with a time length of 0.6 seconds at 5% false alarm level. |
