Live from InterACT: Cutting-Edge Analytics to Predict Risky Driving (DriveCam)
I am attending InterACT San Francisco 2007 this week and blogging live (or nearly live) from the sessions. This session is "Cutting-Edge Analytics to Predict Risky Driving (DriveCam)". DriveCam is privately held and focused on how to help solve some serious problems - 1.2M fatalities, $500B annual cost, 50M injured or disabled. 95% are caused by driver error and there is a direct correlation between specific actions and crashes. Near misses and minor crashes are caused by the same actions as major crashes. Behavior is driven by consequences and, for most people, this is only really true when they see a cop or have their parents in the car. Four different areas:
- Crash recording tools - improves claims processing but does not help prevent crashes
- Driver-focused tools - GPS, speed limit tracking etc but not really preventative
- Crash prevention - lane change or collision warnings but can't tell you if the event was a driver problem or not
- DriveCam - tracking behavior of drivers that causes problem scenarios
The cycle is to identify risk - assess risk - behavior modification - risk reduction - repeat. They use a recording device with an inside and an outside camera/sound that overwrites unless there is some sudden change in the car's movement when it keeps 10 seconds before and 10 seconds after (it keeps all the g-force information with the clip too). The thresholds for sudden change are set by car type.They showed video clips shown that were triggered by a bumpy road (not a problem) and by a sleepy driver (a problem).
When something happens the device makes it clear that it is recording the driver typically reacts immediately - behavior modification. The clips then get reviewed to see what might have caused the problem. The cameras dump data by wifi when the vehicle reaches home base but also using cellular technology now. This transmits data from events immediately and has the potential for real time intervention. DriveCam want to focus in on the riskiest behavior so that the fleet manager to coach the driver - no good telling them "you braked hard a lot today", need to focus on really dangerous behavior. The program drives down claims 30-60%!
Worked with Fair Isaac to help identify the clips that need the most time. Worked to analyze data to help eliminate clips (some vehicles have characteristics that trigger the camera more often for instance), find the risky event triggers and build a foundation for anlaytics. Objectives:
- Maximize the separation of low from high risk
- Maximize hard acceleration, hard braking, hard cornering tracking to high risk
- Cannot miss any collisions
Sampled the clips and extracted the data from g-force etc. Developed new characteristics using the data and identified the data that was inherently predictived. Used Model Builder for Decision Trees to build a segmentation tree. There is a lot of data to be analyzed in parallel. Found some data, like audio, was very predictive but not as usable as it could be (the audio was over-saturated) and fed that back to DriveCam. Highly predictive things turned out to be things like maximum g-force, variance in g-force or length of time the g-force was out of the ordinary. High risk events were often short and collisions particularly abrupt. Low risk patterns often longer in duration. Hard cornering and hard braking also had distinct patterns.
Got 100% on collision detection, 83% correct identification of rough road and 90%+ for hard braking, hard cornering or hard acceleration. This is before adding more data like GPS, speed etc. (DriveCam is working to add mapping partner to be able to track speed limit at time of event as well as integrating with the vehicle to get absolute speed). Also working on a driver risk score that combines this with standard insurance data.
You can find the full set of posts from InterACT in this category.