Think You’re Good Behind the Wheel?
Ever since the widespread release of the automobile, people have been looking into solutions to making roads safer. The focus has been especially on driver behavior that has led to crashes. This consists of acceleration, braking, lane changing, overpassing, and cornering. These behaviors can be summed up into three categories of driving styles: conservative, normal, and aggressive. The ‘conservative’ driving style is commonly associated with a larger braking angle and lengthier changes in speed. The ‘aggressive’ style is often synonymous with faster speeds and acceleration, in addition to frequent lane changes and harder braking, which are often seen as bad behaviors. Lastly, the ‘normal’ consists of more stable actions as the middle ground between ‘conservative’ and ‘aggressive’. It is also seen as the optimized version for better fuel efficiency, safety, and gas emissions.
Previous approaches
Although there have been many proposed methods for identifying and classifying driver behaviors, they all seem to have flaws in one way or another, as either their method of research is subjective, or they don’t consider other factors that can contribute to certain driver behaviors like heart rate, stress, gender, age, education, and job. Other approaches have consisted of probabilistic and machine-learning tools that would process and classify various driving styles. However, the problem is that it often requires longer training times when dealing with increasing numbers of states. In addition, it becomes difficult to classify driving styles as simply ‘safe’ or ‘unsafe’ when the population of drivers increases. To address these shortcomings, later approaches have utilized machine learning algorithms like Neural Networks, K-means, and Decision trees.
Based on the existing literature, there have been many approaches that have shown promising results in classifying driving styles. These approaches collect data through various sources such as onboard sensors, instrumented vehicles, questionnaires, and mobile apps. However, these approaches tend to focus only on one specific behavior. This does not suffice because drivers exhibit multiple behaviors in a single drive, so there needs to be a broader solution that will classify all the behaviors at once. Another key feature that is missing from these approaches is statements of legal implications of certain driving styles. This is an important feature because it would provide drivers with appropriate feedback in addition to assisting with upholding accountability systems when drivers are using company vehicles.
Driving styles from a legal perspective
The authors propose a framework to assess the legal implications of driving styles and offer drivers appropriate feedback. This includes six driving states:
- Normal: Driving adheres to ongoing road traffic policies, no additional feedback
- Restitution: Driving will move to this state if any road traffic policy is violated based on the driving actions. The violation is assessed and classified, and a cumulative penalty is calculated based on when, where, and how many times the violation occurred, which suggests the likelihood that the driver will be a repeat offender.
- Demerit: If the driver commits a more severe offense, they are hit with a demerit penalty. Their driving license could be cancelled or suspended if they cross a certain threshold of demerit points over a certain time. From this state, the driver goes to either ‘Incapacitation’ or ‘Rehabilitation’. However, they can alternatively go back to ‘Normal’ if they don’t commit any more infractions over the set period.
- Incapacitation: This state prevents the driver from driving until after a certain period per the rules defined by the policymakers.
- Rehabilitation: Personalized policies are applied based on driving performance and history. The driving behavior will be assessed, and the driver has a chance to return to the ‘Normal’ state according to the results of the assessment.
- Rewarding: Drivers who have excellent driving behaviors according to the policies during a set period will be given a reward, which might be something like an insurance rebate.
Proposed approach
In this academic study, the writers propose a divide-and-conquer approach that would intelligently recognize and manage various driving styles. As mentioned previously, we need a comprehensive solution that would address all driving behaviors concurrently. Data about each driving behavior would be individually processed to produce a set of decisions, recognize the related driving style, generate a driving score, and formulate appropriate recommendations.
This approach would also help to focus on the most related driving behaviors at the right time and location matched to the right driver. For instance, it would make more sense to focus on braking style rather than the car-following style in a roundabout, whereas car-following style would be more relevant than acceleration style when there is dense traffic. Several recommendation modules would be used to accurately order driving behaviors by priority and share the right information with neighboring vehicles.
MAS and IFAF
This approach depends on a four-layer Multi-Agent System (MAS) architecture that would allow for the management of synchronous driving behaviors in addition to related data, relevance, scores, feedback, and changing contexts. The four layers of the solution consist of Injection, Filtering, Action, and Feedback. These layers integrate intelligent software agents that will allow for the simultaneous management of individual driving behaviors in recognizing their related driving styles and generating relevant feedback.
Proof of Concept
To show how the solution works, the researchers collected driving data through a mobile app called AWARIDE. They tested the app in multiple locations for several drives. This allowed for the collection of driving data from individual drivers, and the parameters collected included speed, acceleration, cornering, wind, temperature, and humidity. The latitude, longitude, and time were also collected with each reading. Here is an example of the acceleration and braking data from two commutes, and then the classifications of both driving styles:
What we can see is that when there is an acceleration event that passes a certain threshold, then the driving style moves toward ‘Aggressive’. The goal here is not so much to compare the two commutes, but rather identify driving styles and the transitions that might happen between them.
Leveraging ubiquitous computing for the classification of driving styles offers a comprehensive approach by integrating sensors in vehicles and smartphones. This enables the collection of diverse data, which, when subjected to feature extraction and machine learning models, allows for the real-time classification of driving behaviors. The incorporation of feedback mechanisms, cloud integration, and user customization enhances the system’s adaptability and utility. However, it is crucial to prioritize privacy considerations and compliance with regulations to ensure responsible data handling. The continuous improvement of the model based on user feedback ensures the system remains dynamic and aligns with individual preferences. Ultimately, the application of ubiquitous computing in this context holds promise for promoting safer and more personalized driving experiences.
Link to academic study: https://link.springer.com/article/10.1007/s00779-023-01740-1\
