Driver Profiling for Autonomous Vehicles

Backstory: ⅓ of Americans currently say they would never buy an autonomous car, regardless of the promised safety and convenience. So it really is about trust now, not just technology, for wider adaptation of these cars.

Impact: Create contextual, semi custom displays by transforming data into displayed information and in turn, understanding into trust.


My Role: Design, Evaluation, Software Development


Timeline: August 2017 - Present

Team Members: Ishaani | Jason J Paul

Impact: Create contextual, semi custom displays by transforming data into displayed information and in turn, understanding into trust.


My Role: Design, Evaluation, Software Development


Timeline: August 2017 - Present

Team Members: Ishaani | Jason J Paul

Research Question


"Given a single type of driving environment (such as highway driving), do different passengers experience different levels of trust in the self driving car, when using Heads Up Displays (HUD) with information displayed in a way personalized for their driving style."


Go to top of this section

Process


Figure: Iterative Process followed in clockwise direction
Following figure shows process carried out during the project. Rigorous literature review was carried out before creating a hypothesis which led to emergence of our research question. In order to test the hypothesis, three passenger profiles were created categorizing them into Thrill Seeking, Defensive and Transit modes of driving. For each profile, four metrics were chosen: Speed, Route being taken, Traffic pattern identification and Lane Keeping. Finally to evaluate these characteristics assigned, card sorting was used as the method of evaluation.


Go to top of this section

Research


What are the external and preexisting trust factors and trust models?

TRUST FACTORS

In order to fully give control to someone/something else, it is important that the person letting go of the control trusts the other person/system. Hence, it is important for autonomous vehicles to develop this trust in their customers to improve their user experience and confidence in the car.
"It’s not the technology. It’s user acceptance that’s holding us up right now"
-Professor [of engineering practice at USC, Jeffrey] Miller
Trust is built over time through a consistent relationship
-MIT Technology Review: Lazy Humans Shaped Google’s New Autonomous Car
People can get comfortable with "their" self driving car
-MIT Technology Review: Lazy Humans Shaped Google’s New Autonomous Car

LEVELS OF AUTONOMOUS VEHICLES

We’re interested in trust models and displays as they apply to Level 5 Full Autonomy cars.
Figure: Levels of Autonomous Vehicle

ADAPTING TRUST MODEL

Adapting the trust model into Head Level Displays, we re-define the factors involved in the trust model
Cognitive Trust: is a customer’s confidence or willingness to rely on a service provider’s competence and reliability
Affective Trust: is the confidence one places in a partner on the basis of feelings generated by the level of care and concern the partner demonstrates
Service Provider Expertise: Is this driver like me, given that I am an expert at driving? Seeing what the system is seeing helps humans understand their “driver”
Similarity: Is this driver like me? Does it care about me? Does it share my values?
Figure: A model of customer trust in service providers

Go to top of this section

Hypothesis


From our research, we hypothesize that
  • Passengers do not care about how the Car AI is actually thinking/perceiving/calculating. Passenger care about whether it’s “paying attention” to the things they would.
  • Displaying different information in different “Heads Up Display Profiles” that map to the styles in which people drive can increase passenger trust of the AI driver system.

Go to top of this section

Design of Driver Profiles


We categorized drivers into three profiles: Thrill Seeking, Defensive and Transit and used four metric: Speed, Route being taken, Traffic pattern identification and Lane Keeping; in which these three profiles would differ in terms of their response and display prominence.
Thrill Seeking: Get me there as ENTERTAININGLY as possible. This profile seeks to cater towards people who are more adventurous.
Defensive Driving: Get me there as SAFELY as possible. The main aim of passenger falling in this profile is to get to their destination following all rules and regulations.
Transit Driving: Get me there as EFFICIENTLY as possible. The main aim of passenger falling in this profile is to get to their destination by maximizing output using minimum number of resources.
The color denotes emphasis level given to each area of driving metrics, and the shape indicates the whether the information needs to be acted upon. The reason this second dimension is important to differentiating the profiles and matching the display of the metrics in that area with an appropriate level of context.
Figure: Different Profiles perceive and react to various metrics

TRANSIT

Figure: Transit profile visuals

DEFENSIVE

Figure: Defensive profile visuals

THRILL SEEKING

Figure: Thrill Seeking profile visuals



Go to top of this section

Evaluation


How well do the profiles represent driver behaviour?

We use In Person Computerized Between Subjects Closed Card Sorting technique to evaluate our hypothesis: Drivers can be divided into three driving profiles: Thrill Seeking, Defensive and Transit.
Users will be asked to categorize the visuals into the following sets
  • Thrill Seeking vs Non- Thrill Seeking
  • Defensive Vs Non - defensive
  • Transit Vs Non - transit
*The evaluation stage is still in progress and we are currently in the phase of developing custom card sorting program to support the desired functionality

Go to top of this section