On Control Transitions in Autonomous Driving: A Framework and Analysis for Characterizing Scene Complexity

Nachiket Deo, Nasha Meoli, Akshay Rangesh, Mohan Trivedi; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

Abstract


'Take-overs' are safety critical events in conditionally autonomous vehicles. These are cases where vehicle control is transferred from the autonomous system to a human driver during failure modes of the system. Safe take-overs depend on two key factors; the readiness of the driver, and the complexity of the scene. While prior work has addressed driver readiness estimation, scene complexity estimation for control transitions remains an unexplored topic. In this paper, we focus on characterizing the complexity of driving scenes as perceived by human drivers during takeover events. To this end, we collect naturalistic driving data using a conditionally autonomous vehicle, equipped with cameras and LiDAR sensors. We mine a diverse set of scenarios using the LiDAR point cloud statistics. We then collect take-over complexity ratings in these scenarios assigned by raters with varying degrees of driving experience. We present an analysis of inter-rater agreement, and the average rated complexity conditioned on features of the surrounding environment, detected agents around the ego-vehicle, and ego-vehicle actions and motion states.

Related Material


[pdf]
[bibtex]
@InProceedings{Deo_2019_ICCV,
author = {Deo, Nachiket and Meoli, Nasha and Rangesh, Akshay and Trivedi, Mohan},
title = {On Control Transitions in Autonomous Driving: A Framework and Analysis for Characterizing Scene Complexity},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
month = {Oct},
year = {2019}
}