첨단 운전자 주의력 분산 경고(ADDW) 시스템

The European Commission has released a report on ADDW systems, in response to the Driver Distraction and Drowsiness Recognition safety measure forming part of the European General Safety Regulation. ADDW is a system that recognizes driver attention levels and warns if the driver becomes distracted.

An ADDW system is comprised of two parts:

1. Tier-1 suppliers provide technology to determine or estimate whether the driver is looking forward at the road ahead or has their attention diverted somewhere else, by monitoring a driver’s eyes, face and/or head whilst driving using a driver-facing camera embedded in the steering wheel or instrument cluster.

2. Car manufacturers (OEMs) are responsible for taking this piece of technology and integrating it into the vehicle, developing algorithms to detect distraction and non- distraction events, optimising the interaction with the vehicle and driver, and interacting with the driver when needed. Many OEMs are still developing their ADDW systems, and several do not yet have a system in development.

Two main issues which need to be overcome in order for the systems to be deemed effective and implemented into vehicles, are the reliability and robustness of the systems.

Reliability - manufacturers expressed that their systems are displaying an undesirable number of:

• False positive alerts, which negatively impacts customer experience and trust in the system (i.e. will ignore or turn it off), as well as potentially causing distraction instead of preventing it; and

• False negative alerts which defeats the purpose of the system and negatively effects driver’s trust in the system (i.e. not alerting a driver when they are distracted).

One of the main challenges causing false positive and false negative alerts surrounds the difficulty in determining whether or not someone is visually distracted, e.g., they may have removed their eyes to perform an activity related to the driving task such as checking mirrors and blind spots, reading road signs, scanning the windscreen and gazing out of side windows to assess the environment or traffic situation. Some systems have attempted to address this by defining multiple areas of interest (AoI) such as the mirrors, instrument cluster and Forward field of view (FoV) and have coded these AoI as relevant for driving, meaning that when a driver gazes at one of these areas, the system will not deem the driver as being visually distracted. Stakeholders stated that specifying multiple AoI and making use of secondary inputs made their systems more robust and reliable. For the majority of systems, the size of the Forward FoV was dependent on vehicle speed, where the Forward FoV was smaller at higher speeds than at lower speeds. Some stakeholders stated that one of the reasons for this was due to the driver attention zone being wider in urban driving environment compared to motorway driving environments. This means that some systems would allow more head movement in urban environments (i.e. wider Forward FoV), but shorter time duration before the warning, whereas, in a motorway environment, the system would allow less head movement (i.e. narrower FoV), but a longer time before a warning would be considered.

A few systems have incorporated secondary metrics such as indicators, ADAS and the external environment into their algorithm to understand the driver’s intentions, and hence assist in determining whether a gaze is related to the driving task or not. For example, a driver gazing frequently or for a long duration towards the offside window may be interpreted as a distraction event (e.g. engaging with a passenger). However, the driver may also be assessing the traffic situation when stopped at a junction waiting to turn onto a busy perpendicular road (i.e. waiting for a gap). In this situation, information about the turn signals and vehicle speed will assist the system in determining the driver’s intentions and infer that the driver is not distracted. This method is a recent development and manufacturers require more time to refine and validate the method, where some manufacturers are not yet at this stage of development.

Another factor influencing the number of false positive alerts, which stakeholders are still trying to understand and incorporate into their systems, is the breadth and limitations of human vision. Human vision is divided into three ranges: foveal range, extrafoveal range and peripheral vision range. Vision is most detailed in the foveal range, which extends approximately one eccentricity angle for the line of sight, followed by the extrafoveal range, extending about 30 degrees eccentricity angle, whereas the last range, the peripheral visual range, extends from 30 degrees eccentricity angle up to 100°-110° horizontally (away from the nose towards the ear), 60° up and 70°-75° down. For both eyes, the combined visual field is 130°-135° vertically and 200°-220° horizontally.

Another factor highlighted by stakeholders as negatively impacting the reliability and robustness of ADDW systems is the difficulty in establishing appropriate distraction thresholds. When a driver is visually distracted and engaged in a secondary task, they tend to either glance away from the road for an extended period (Type 1) or, most commonly, shift their gaze between the two tasks frequently until the secondary task is completed (Type 2):

1. Type 1 - The most widely recognised guidance in the literature was that a single glance should not exceed a two-second duration.

2. Type 2 - For Type 2 distraction behavior, the following measurements were identified:

• Percentage Road Centre (PRC): percentage of time within 1 minute that the gaze falls within a road centre area of 8 ̊ radius from road center.

• Total Eyes-Off Road (TEOR): the summation of all glance durations to areas of interest other than the road scene ahead, in particular, TEOR glance duration greater than two seconds in a six-second window.

• Total Glance Duration: summation of all glance durations to an area of interest, in particular, a distracting task should be completed within 15 or 20 seconds of total glance duration.

• Glance Frequency: number of glances to a target within a pre-defined time period, or during a predefined task, where each glance is separated by at least one glance to a different target.

It was suggested that the sensitivity of the threshold is dependent on the driver’s visual awareness of the road (i.e. sensitivity and threshold dependent on the eccentricity angle from the target to the center of the road). Moreover, stakeholders suggested that the sensitivity of the threshold is also dependent on the driving context (i.e. time of day, road type, traffic, drowsiness level etc.).

The minimum performance requirements for the human machine interface (HMI) for ADDW systems need to be established and included in the regulation. If the system does not interact with the driver in an appropriate manner or does not inform the driver that they are distracted effectively, the driver may ignore the alert, potentially resulting in collisions. Additionally, if the alert is not designed appropriately, or if the system is not sufficiently robust (high false positive rate), it may cause distraction or driver annoyance.

ADDW systems shall be designed in such a way that they shall only continuously record and retain data necessary for the system to function and operate within the closed loop system. Furthermore, this data shall not be accessible or made available to any third parties and shall only be held for the length of time for which it holds direct relevance to assessing the driver’s current visual attentive state.

We at Neonode fully agree with the message in the report and encourage everyone to read it thoroughly. Neonode has solutions for the corner cases described in the report and a capable toolbox for configuring a complete DMS system - from start to finish.