United States Department of Transportation
i
Machine Learning, Human Factors and Security Analysis for the Remote Command of Driving: An MCity Pilot
-
2019-12-01
-
[PDF-617.82 KB]
Details:
-
Creators:
-
Corporate Creators:
-
Corporate Contributors:
-
Subject/TRT Terms:
-
Publication/ Report Number:
-
Resource Type:
-
Geographical Coverage:
-
Edition:Final Report (Sept 2018 – Dec 2019)
-
Corporate Publisher:
-
Abstract:Both human drivers and autonomous vehicles are able to drive relatively well in frequently encountered settings, but fail in exceptional cases. These exceptional cases often arise suddenly, leaving human drivers with a few seconds at best to react— exactly the setting that people perform worst in. Autonomous systems also fail in exceptional cases, because ambiguous situations preceding crashes are not effectively captured in training datasets. This work introduces new methods for leveraging groups of people to provide on-demand assistance by coordinating responses and using collective answer distributions to generate responses to ambiguous scenarios using minimal time and effort. Unlike prior approaches, the authors introduce collective workflows that enable groups of people to significantly outperform any of the constituent individuals in terms of time and accuracy. First, the authors examine the latency and accuracy of crowd workers in a future state prediction task in visual driving scenes, and find that more than 50% of workers could provide accurate answers within one second. The authors found that using crowd predictions is a viable approach for determining critical future states to inform rapid decision making. Additionally, the authors characterize different estimation techniques that can be used to efficiently create collective answer distributions from crowd workers for visual tasks containing ambiguity. Surprisingly, the authors discovered that the most fine-grained and time-consuming methods were not the most accurate. Instead, having annotators choose all relevant responses they thought other annotators would select led to more accurate aggregate outcomes. This approach reduced human time required by 21.4% while maintaining the same level of accuracy as the baseline approach. These research results can inform the development of hybrid intelligence systems that accurately and rapidly address sudden and rare critical events, even when they are ambiguous or subjective.
-
Format:
-
Funding:
-
Collection(s):
-
Main Document Checksum:
-
Download URL:
-
File Type:
