ABSTRACT:
The process of driving a car involves a cognitive load that varies over time. Additional load comes from secondary factors not directly associated with the driving process, including navigation devices, entertainment systems and the car’s own warnings.
In this paper, we present a framework for intelligent scheduling of in-car notifications based on the driver’s estimated cognitive load. As the single channel for communication, it reschedules the notifications using a priority queue, and relays them to the driver based on the urgency of the notification and the overall estimated cognitive load being experienced by the driver at any given moment.
We evaluate our system using a dataset collected from a car’s CAN bus during multiple on-road trials and show that our proposed approach reduces the number of simultaneous calls on the driver’s attention during the driving task. We also demonstrate that our intelligent scheduling significantly reduces the maximum cognitive load experienced by the driver and the frequency with which high loads occur.
DATA COLLECTION
The dataset we use is collected in real driving scenarios. Eight participants took an instrumented vehicle around the same 90-minute route, chosen to include urban, rural and highway sections to get diverse driving conditions. Activity inside and outside the car was captured by several cameras, and data from the car’s CAN bus was recorded, including pedal pressures, the use of controls such as indicators, and steering wheel angle.
Partitioning Of The Message Queue into Five Smaller Independent Queues According to Importance Classification.
SYSTEM DESIGN
In the Apollo space missions, all communications to the astronauts from mission control went through a single individual, to avoid the possible confusion of multiple simultaneous communications. This individual, designated the ‘Capsule Communicator’ or CAPCOM, was typically a trained astronaut who had some understanding of what the crew was going through.
Message Classification and Priority Queue:
To classify the different messages and define the mechanics of the queue, we associate two different values with each message: Message Importance and Message Priority.
We define the criteria for a message to be placed in one of the above categories as follows:
- Critical messages are those requiring a prompt response from the driver. Relayed immediately, they generally convey information about safety risks such as seatbelts not being in use, or urgent navigation instructions.
- Very High importance messages alert the driver to situations where action is required in the short term. These messages include fuel and vehicle fault warnings.
- High importance messages convey information which does not require immediate action, such as traffic warnings.
- Medium importance messages contain notifications of future events, such as a service due date.
- Low importance messages are purely for the driver’s entertainment. They include social media and interesting landmark notifications.
Cognitive Load Estimation:
The cognitive load (CL) resulting from the driving task was estimated using the CAN data from the vehicle. Although correlations have been presented, very little work has been done previously to quantify the factors which affect thedriver’s CL. We have defined and modelled five factors believed to affect the driver’s CL (summerised in Figure 3). Each of these factors was defined on a scale between 0 and 1 and then scaled by a constant factor based on estimates of how large an effect it would have on the driver’s CL.
Modeling of Cognitive Load (Cl) as a Function of Different Car Parameters.
EVALUATION
To evaluate our framework performance when applied to data captured in real driving scenarios, we created simulations where a range of events/notifications were generated and injected into the recorded driving data as if they had been generated by in-car systems.
Our intelligent scheduling framework takes these, queues them up, prioritises them and delivers them at the appropriate time according to the approach described above. To assess the effectiveness of our system, we modelled the additional cognitive load resulting from the messages being fed to the driver.
Case Study Analysis:
Given our models of the baseline load caused by driving and the additional load resulting from notifications, we can add these together to get the approximate overall load experienced by the driver at any given moment.
It may be argued that the load resulting from context-switching in the case of multiple simultaneous tasks is greater than simply the sum of those tasks when experienced individually, but we have not assumed any multi-tasking overhead in our model.
Overall Cognitive Load Analysis:
For a broader evaluation of the effect of the scheduler on the overall cognitive load of the driver, we analysed eight CAN data files from the dataset described in ‘Data collection’ above, one for each of 8 different participants driving our full route ( 10 hrs of driving in total).
CONCLUSIONS
In this paper, we have presented an intelligent scheduling system for in-car notifications which aims to reduce the overall cognitive load experienced by drivers from in vehicle systems. We described the two components of our system, namely: message classification using priority queues, and cognitive load estimation. We evaluated our approach using CAN-bus data collected from a dataset of real-world recordings.
We demonstrated that our scheduling approach significantly reduces the peak cognitive load experienced by the driver and the frequency with which high loads occur. Thus, our system smooths out the cognitive load on the driver by presenting less urgent messages at a time when the driver is not occupied with difficult driving situations.
FUTURE WORK
Future work includes comparing our system’s rescheduling of messages to the timing that human passengers would use based on their judgment of the driver’s cognitive load. We also plan to incorporate additional modalities for cognitive load estimation, such as gaze and head pose.
Finally, the scheduling could also be made dependent on external sources such as the proximity of other vehicles, and events likely to be encountered in the future, such as approaching traffic congestion or upcoming navigation directions.
Source: University of Cambridge