Seeing Machines Lauched by Volvo, ANU
IVsource.net
16 October 2000

A research collaboration between Volvo Technology and Development and the Australian National University has resulted in the formation of this new company, whose products are geared towards detecting fatigue and inattention, interactive multimedia, and human performance measurement using newly developed techniques.

Image processing continues to unlock new doors between the human and machine world.  As Seeing Machines puts it, "Our new 'Face Interface' technology takes you into a new realm of simplistic and natural interaction with machines. Our cutting-edge face and gaze tracking technology opens the door to a world that is simpler, safer and smarter."

Noting that your face is a natural interface mechanism, they assert that their technology can interpret a person's face and measure intention ... or inattention.  The technology has obvious potential for driver monitoring, if used to detect inattention to key driving tasks (e.g., driver distraction due to telematics interaction) or to detect fatigue or drowsiness.

And because the system not only interprets facial activity but also creates a "virtual face," mimicking a human in real-time, the company is also eyeing new applications for smart computer interfaces and new forms of interactive multimedia in entertainment and business applications.  (Ya gotta see it to really get the picture -- their demo video, accessible at www.seeingmachines.com, is worth the download time! -- Ed.)

A New Approach to Eye-Gaze Tracking

According to the company's website, Seeing Machines was founded to commercially develop a new type of human-machine interface, developed by the Research School of Information Sciences and Engineering at the Australian National University (ANU).  The research was sponsored and Volvo Technology and Development as part of their ongoing quest to find and develop promising new technologies in human performance measurement.  For example: Volvo wants to test the safety of new cabin designs by determining whether particular cabin layouts or features (such as televisions) are likely to confuse or distract the driver.  Fundamentally, this means they need to know where the driver is looking at any given time.  After studying and testing the state of existing eye-gaze-tracking technology they concluded that their requirements for reliability and robustness simply weren't met by any other available system.

The company says they invested eight man-years of research effort to develop this technology, which effectively gives machines the ability to 'see' people ... enabling them to automatically detect and track human faces, in real-time, using only stereo-video information.  The technology is also capable of performing precise and reliable facial measurement, and is invariant to head-motion and rotation.  A by-product of the approach is the ability to accurately and reliably determine a person's direction of gaze, they say.

Specifically, their system is described as being able to determine:

• Human head position and orientation, in three dimensions -- the head must be within a working volume with approximate dimensions of a meter high, a meter wide, and 1 meter in depth (therefore enough to completely capture the range of possible head positions when sitting in a chair or a driver's seat)

• Eye-gaze direction, in three dimensions -- the system can detect if one or both eyes are obscured, and provides a gaze estimate that degrades gracefully with increasingly difficult measurement conditions.  The measuremen works even with drivers wearing eyeglasses or sunglasses.

• Eye closure and blink rate -- these measurements are only available when the eyes are visible.  They can provide the necessary information to detect fatigue and boredom.

• The location of certain flexible features on the face -- for example, the eyebrows and corners of the mouth can be tracked to determine facial expression.

Stereo Vision Does the Trick

Seeing Machines' technology operates through passive analysis of stereo video information.

The software is able to locate and identify a human face and then track it, by locating the 3D positions of specific facial features.  If these features become obscured then they are dynamically replaced with new features.  The system is thus able to keep tracking the face when the head is rapidly turned, or when the face is partially obscured.

Why two cameras?  Research into monocular face tracking has revealed that while it is theoretically possible to make 3D measurements of head pose from a single viewpoint, the practical issues of finite image resolution and processing power place tight restrictions on the precision and accuracy of potential techniques.  According to Seeing Machines, the emergence of lower-cost CMOS camera technology has tipped the cost-benefit equation towards stereo.  In fact, they say it is likely that as video capture and processing power costs continue to fall, the balance between cost, complexity, and performance will shift towards systems with even more cameras.

Dealing with the Practicalities Inside a Vehicle

Other approaches to measure head pose or eye gaze include the aforementioned monocular image processing and 'classic' eye-tracking.  For some time, eye-tracking has been effected with one or more infrared emitters pointed at the face, and a dash-mounted camera whose image is processed to detect the infrared energy reflecting off the retinas.

However, both techniques are limited to some degree by the following:

• Head motion must be either slow, or contain no depth variation and no rotation;

• Tracking cannot be done in three dimensions;  and

• Failure, false measurement and noise sensitivity under conditions of head motion, perspective distortion, feature occlusion, and lighting fluctuation.

The company says that, because the technology has been developed principally for detecting driver fatigue and inattention, a particular emphasis has been placed on robustness and reliability, and so the technology does not suffer from any of the problems mentioned above.

So are eye trackers based on infrared emitters/reflectance a thing of the past?  It's too early to say, but those seeking to apply traditional eye tracking to automotive applications have had difficulty dealing with head movement, various issues with reflectance of the retina, and lighting conditions.  It is also true that the accuracy of the typical eye tracker is far more than needed to determine if a driver is looking at a navigation screen, side mirror, or straight ahead.

And how would consumers react?  Automakers have been leery to consider installing systems which fire an infrared beam continuously at drivers -- considered safe by industry standards, but potentially a major problem in terms of consumer perceptions.  If the job can be done passively, all the better.  Of course, some would say that drivers don't want their mug continuously monitored under any circumstances -- but this is the type of factor handled by marketing and emphasizing the benefits of the system.

This sifting-out of technology approaches will continue for a while yet.  According to Tim Edwards, principal engineer for the company, they are actively working to build a system for fatigue management for the transport industry as a whole. "Such systems are unlikely to be on the market for a few years yet. One of the major hurdles for everyone trying to address this issue will be reducing the false alarm rate whilst ensuring reliable alarms for real fatigue safety events. A high false alarm rate will cause such systems to be ignored or disabled," he says.  Thus, they are continuing to work with Volvo TD to develop the sensing means to reliably detect fatigue, situations of high cognitive load, and other hazardous situations.

[Top]

... contact Tim Edwards, Principal Engineer, at tim@seeingmachines.com -- or see their web site at www.seeingmachines.com.  

[Top]