Multi-Source Feature Extraction and Visualization for Driving Behavior Analysis

Abstract

The high-dimensional multi-source driving behavior data is too difficult for people to understand. And the low-dimensional data visualization is a more intuitive way to represent the driving behavior data. However, the previous driving behavior visualization methods only evaluate the passive vehicle's ego behavior data, and omitted the active driver's attention. To solve this problem, a driving behavior feature extraction and visualization method based on multi-source data fusion is proposed. Firstly, the driver's head pose data (including yaw, pitch and roll) and vehicle data (including speed, acceleration and engine speed data) are collected through in-car cameras and vehicle OBD (On Board Diagnostics) interface, respectively. All these data are normalized and from which time series data are extracted by sliding windows method. Then, in order to take advantage of these multi-source data, FastICA is used to extract 3D hidden features of driving behavior. Finally, the 3D hidden feature is mapped to the RGB color space. A colored trajectory is produced by placing the colors in the corresponding GPS data. Experimental results demonstrate that the proposed method has a higher average F-measure value than other baseline methods. The colored driving behavior trajectory can help people better distinguish different driving behavior.