Automatic Unusual Driving Event Identification For Dependable Self-Driving

Authors:
Hongyu Li Rutgers University
Hairong Wang Rutgers University
Luyang Liu Rutgers University
Marco Gruteser Rutgers University

Introduction:

This paper introduces techniques to automatically detect driving corner cases from dashcam video and inertial sensors. We evaluate the system based on more than 120 hours real road driving data.It shows 82% accuracy improvement versus strawman solutions for sudden reaction detection and above 71% accuracy for rare visual views identification.

Abstract:

This paper introduces techniques to automatically detect driving corner cases from dashcam video and inertial sensors. Developing robust driver assistance and automated driving technologies requires an understanding of not just common highway and city trafic situations but also a plethora of corner cases that may be encountered in billions of miles of driving. Current approaches seek to collect such a catalog of corner cases by driving millions of miles with self-driving prototypes. In contrast, this paper introduces a low-cost yet scalable solution to collect such events from any dashcam-equipped vehicle to take advantage of the billions of miles that humans already drive. It detects unusual events through inertial sensing of sudden human driver reactions and rare visual events through a trained autoencoder deep neural network. We evaluate the system based on more than 120 hours real road driving data. It shows 82% accuracy improvement versus strawman solutions for sudden reaction detection and above 71% accuracy for rare visual views identification. The detection results proved useful for re-training and improving a self-steering algorithm on more complex situations. In terms of computational eficiency, the Android prototype achieves 17Hz frame rate (Nexus 5X).

You may want to know: