Driving Events and Scores
Our platform estimates various risky events and features, which are combined into scores. These scores are used to evaluate driving behavior and safety. The features and scores available are as follows:
Phone Usage Events
Calling: Events reliably detected by our SDK when a call is made during driving.
Screen Use: Events reliably detected by our SDK when the phone screen is used while driving.
Phone Handling: Events predicted using ML models to detect active phone handling while driving.
Harsh Events
Braking
Negative longitudinal acceleration surpassing 3 m/s2 (configurable).
This refers to a sudden decrease in speed or a "hard brake" when driving a vehicle. It's like stepping on the brakes abruptly, causing the car to slow down quickly.
Example: Imagine you're driving, and suddenly you have to hit the brakes very hard to avoid an obstacle on the road. This action creates a negative longitudinal acceleration.
Accelerating
Positive longitudinal acceleration surpassing 2.25 m/s2 (configurable).
This refers to a sudden increase in speed or a "hard acceleration" when driving a vehicle. It's like stepping on the gas pedal quickly, causing the car to speed up rapidly.
Example: Picture yourself at a traffic light, and when it turns green, you push the gas pedal down very hard to quickly accelerate and merge into traffic. This action creates a positive longitudinal acceleration.
Turning
Lateral acceleration surpassing 3.5m/s2 (configurable).
This refers to the force that pushes a car to the side when making a sharp turn. It's the feeling of being pushed to the side in your seat as the car takes a curve quickly.
Example: Imagine you're driving on a winding road, and you take a sharp turn at a high speed. You feel the force pushing you to the side as the car turns, creating a lateral acceleration.
Legal Events: Measures adherence to speed limits by detecting when speed surpasses OSM speed limits.
Contextual Data: Time of Day, Distance Driven, Road Type, etc.
Flexible Score Calculation Logic
Our platform has a flexible score calculation logic that combines raw events and applies necessary thresholds and trip-level and population-level normalization. When you start to use our platform you will receive the following scores as defaults. These scores were validated by multiple clients and tuned to be as universal and useful as possible.
Speed Based
Legal: The legal driving score measures how well you adhere to speed limits. The higher your score, the more you respect the speed limits.
Dynamics Based:
Harsh Accelerating: Measures how often you accelerate hard. Every hard acceleration will be penalized by subtracting a percentage of your score.
Harsh Braking: Measures how often you need to brake hard. Every hard brake will be penalized by subtracting a percentage of your score.
Harsh Turning: Measures how often you turn hard. Every hard turn will be penalized by subtracting a percentage of your score.
Smooth: The smooth driving score measures how smoothly you drive. High accelerations, heavy braking, and heavy turning result in a lower score. Scores are normalized with respect to a wide population. The higher your score, the smoother you drive.
Harsh Events: Combination of HarshAccelerating and HarshBraking scores. The hard brakes and accelerations are also normalized by the total number of events.
Anticipative: The anticipative driving score measures how well you anticipate turns. Hard accelerations before or hard braking during a turn result in a lower score. The higher your score, the more anticipative you drive.
Phone-Usage Based
Attention: A combined score of calling events (weighted differently if the phone is mounted) and active phone handling.
Focus: The proportion of time (percentage) the user is focused while driving, being focused means: not using the phone, which is detected through active phone handling.
Call While Moving: The proportion of time (percentage) the user is on a call while driving their car at higher speeds.
OverallSafety: Flexible way to combine the sub-scores. By default, it is defined as a combination of Legal, Attention, and Smooth scores. If the average overall score for a user is less than 0.7 they can be considered a bad driver.
The above scores are returned for each trip. We found that for engagement purposes (user-facing score) addition smoothing is useful. This ensures that a single bad trip does not influence user score too much to avoid demotivation while at the same time allowing users to see the impact of their behavior change as fast as possible. For this reason, we also have adaptive scores, which take into account user score history.
In addition, we aggregate scores per user at fixed intervals such as daily, weekly, monthly and quarterly. These scores are available as Time Aggregated Scores in the backend.
Frequently Asked Questions (FAQs):
What scores are available on-device?
Currently, we have a subset of scores available on-device (Smooth and Focus scores). We are working on bringing more scores on devices in the near future. Also, no aggregation happens on the device i.e., time aggregated scores are available only in the backend and not on the device.
Which driving events can be accessed directly on the device?
Call events, phone handling events, screen events and harsh events. Note that we do not expose the type of harsh event such as braking, accelerating and turning in the on-device harsh event model. Why might scores be missing or have -1 as the value?
To reliably calculate the scores we ensure that the quality of the sensor signal and duration of the trip is sufficient. Usually, scores will be missing only for really short trips in duration (< 3 min) and/or distance (< 2 km or 1.25 mi). Longer trips might have missing scores due to bad quality sensor data (phone was in flight mode, excessive noise in the signal, etc.), but these are rare occurrences. A score of -1 indicates that we were unable to compute the score due to the reasons mentioned above.
Why was phone usage falsely detected during a trip?
Our algorithm that detects phone usage while driving might sometimes detect phone usage when the phone is moving in the pocket while driving, or when it hits a sequence of speed bumps. We are using ML to look for active phone handling (moving your phone while it’s in your hand).
Last updated