https://ieeexplore.ieee.org/abstract/document/9805363

Authors

Alex Richardson, Kate Sanborn, Jonathan Sprinkle

Abstract

Data sets that contain dash camera and sensor data are essential to the development of autonomous vehicles. Many current methods of constructing these data sets are manually intensive, expensive, and difficult to scale. This paper discusses a method of automating the development of a naturalistic driving data sets that include dash camera footage labelled with information captured from the Controller Area Network (CAN) bus on a vehicle. CAN data can contain IMU data, radar data, etc. First, the dash camera footage and CAN bus data are paired and synchronized using optical flow analysis. Once the footage has been labelled with the telemetric information, one can identify important events in driving behavior by examining the signal conditions. This method is significantly less expensive and more scalable than previous data sets, while providing competitive quality in terms of telemetric data. It could significantly increase the quantity and diversity of driving data sets in the future.