Automatic Impact Detection with Smartphone Telematics – How Crash Detection Can Save Lives

In this episode of the Insurance Telematics Podcast, host Harald is joined by Chief Data Scientist Katharina to discuss the crucial topic of automatic impact detection using only smartphone sensors. The conversation highlights how crash detection can dramatically improve emergency response times and reduce insurance claim costs. With real-world insights and technical explanations, the episode dives deep into how a combination of heuristics and machine learning helps distinguish real accidents from false positives—all with the aim of democratizing vehicle safety.

Why Crash Detection Matters

Harald shares a personal story about a rollover accident he experienced at night—only discovered because a truck driver happened to pass by. This event underlines the urgency of real-time accident detection, especially when injuries may prevent the driver from calling for help.

Key Benefits of Smartphone-Based Detection

• Immediate emergency chain activation
• Scalable solution for all customers
• Cost-effective and inclusive across insurance portfolios

What Is Automatic Impact Detection?

Katharina defines automatic impact detection as the real-time identification of accidents using only smartphone sensors—primarily the accelerometer. The goal is to detect severe impacts and trigger help without user interaction.

Focus Areas

• Detecting accidents with bodily injury
• Filtering out false positives such as phone drops or sudden braking
• Capturing FNOL (First Notification of Loss) data for insurance purposes

How Smartphone-Based Crash Detection Works

1. Sensor Data Collection

• Accelerometer data sampled at 100 Hz
• Peak detection: sudden spikes followed by inactivity

2. Filtering with Heuristics

Rules like: If movement continues a minute after impact, it’s not a crash. Evaluating speed drops and periods of stationarity post-impact. Pre-filtering before sending data to reduce false alarms.

3. Machine Learning Models

• Trained on labeled real accident data
• Differentiates between crashes and other phone movements
• Incorporates contextual inputs like speed and direction of impact

Ensuring Data Reliability and Emergency Readiness

Immediate Data Upload

Data is sent to the backend within milliseconds post-impact to ensure it’s available even if the phone is damaged during the crash.

Acceleration Severity Index (ASI)

ASI estimates injury likelihood:
ASI > 1: likely injury → emergency alert
ASI < 1: low-risk → optional follow-up

Technical Challenges and Solutions

Phone Not Fixed in the Vehicle

The phone’s varying position requires normalization to match the car’s frame of reference and track orientation changes throughout the trip.

Battery Considerations

High-frequency accelerometer sampling is low-impact on battery. Most battery drain comes from GPS and data transmission.

Real-World Accuracy and Scalability

With 10,000 users per year, the system expects around 300 real accidents. It prioritizes critical crashes and allows insurers to adjust the ASI threshold based on risk tolerance and budget.

Final Takeaways

• Smartphones enable democratized accident detection
• Layered AI approach: fast heuristics + deep learning
• Real-time data capture ensures timely emergency response

Thank you for tuning in to the Insurance Telematics Podcast. Stay safe—and smarter—with crash detection driven by smartphone telematics.

---

Harald
Hello everybody. Welcome back to the Insurance Telematics podcast. Today again with our chief data scientist, Katharina. Welcome.

Katharina
Hi. Thanks for having me.

Harald
Sure. Today, we have a very important topic. We’re talking about automatic impact detection, crash detection. And why is this so important?

I had an accident myself where I was only incidentally found by a truck driver because I had a rollover accident at night. I have seen myself what could have happened if nobody would have found me.

So, this is a personal topic that I’m very emotional about. And I want to tell you why I believe it’s very important for the world.

When you have an accident, time counts. And if we can detect accidents with something like a smartphone, we become more powerful in detecting more accidents. Yes, there are better ways, like installed hardware or dongles. But the point is, we want to democratize this feature.

If you think from an insurance perspective, and you can equip 100% of your customers with something that works about 90% of the time, versus equipping maybe 3 or 4% of your portfolio with something that works 95% of the time, I personally would opt for the more available, low-cost solution that can help many people.

So Katharina, we’re now talking a little bit more about the technical aspect. What is automatic impact detection? What’s our goal with it?

Katharina
Automatic impact detection means we automatically detect a car or vehicle accident based on smartphone sensors only, with the goal to trigger the emergency chain as soon as possible. This promotes increased safety and also helps insurers streamline claims processing and reduce time and costs.

Harald
So our goal is to detect accidents. What kind of accidents are we trying to detect?

Katharina
We’re trying to detect accidents that can cause bodily injury. That is our main focus because smartphones are moving devices, not fixed in cars, and can record other events like falling down. So there’s a lot of noise. We focus on accidents with bodily injury.

Harald
What we’re doing with smartphone-detected accidents is saving these mini crashes for FNOL use later in claims management, but we’re trying to forward only accidents likely to involve bodily injury.

Katharina
Exactly. If I may, I would walk you through the process step by step. We said it’s smartphone-only, so all data collection, processing, and interpretation live directly on the smartphone. At a low level, we listen to accelerometer data, sampled at 100 Hz. We listen for peaks—an accident manifests as a high peak followed by nothing for some time.

Harald
A quick explanation for non-technical listeners: acceleration measures gravity across multiple axes, helping us understand object movement in 3D space. 100 Hz means we collect 100 data points per second.

Katharina
Exactly. We listen for peaks. Once we detect a peak, the challenge is to determine if it’s a car accident or caused by something else—like the phone falling. We have a filter logic made of heuristics and machine learning.

Harald
Heuristics are hardcoded rules, like if the car keeps driving a minute after an impact, it’s likely not an accident.

Katharina
Exactly. There are also external sensors like Bluetooth tags or dongles, but we’re targeting smartphones. The difficulty isn’t detecting accidents but eliminating false positives.

Harald
How do we do that?

Katharina
We’re lucky to have years of experience with hardware-based solutions and labeled data—GPS, accelerometer, etc.—where we know what real accidents look like. We train models using this data.

Harald
Labeled means a human said this was an accident. So we train models with patterns.

Katharina
Exactly. We also include diverse non-accident scenarios like phone interactions or hard braking.

Harald
So we have heuristics and machine learning. Can you give an overview of some rules?

Katharina
One key rule is checking for a period of stationarity after an impact. We analyze a time window—30 seconds before the impact and up to 1 minute after—to understand speed, activity, and whether the phone has stopped.

Harald
Right. Speed decline is important. If you’re going 50, hit an event, and then drop to zero, it’s different than dropping to 40.

Katharina
Exactly.

Harald
We use rules first. Why?

Katharina
Rules are quick, effective, and eliminate most false positives—important since we want to trigger emergency chains quickly.

Harald
What happens if a phone is damaged during an accident?

Katharina
As soon as we detect an impact, we store and send data to the backend immediately—within fractions of a second. We don’t wait for filtering.

Harald
So we ensure data is sent before damage might prevent it. Then filtering and machine learning come after.

Katharina
Exactly.

Harald
Can we quantify accuracy? What’s ASI?

Katharina
ASI is the interpretation layer. Once we confirm an accident, we look at acceleration forces—magnitude, direction, and intensity. ASI helps assess severity. Below 1 means low likelihood of injury; above 1 means potential for injury, requiring emergency action.

Harald
After confirmation, we forward the data to a security center via our platform or API. What are other technical challenges?

Katharina
A major one is the phone not being fixed in a car. That causes two issues: detecting non-accidents and difficulty interpreting severity. We normalize smartphone data to match the car’s reference frame, accounting for orientation changes during the trip.

Harald
So we need to learn the phone’s position relative to the car. But it can move during a trip.

Katharina
Exactly. It’s challenging.

Harald
We increase sampling frequency to 100 Hz. Does that impact battery?

Katharina
Not necessarily. High frequency is needed for accurate peak detection. The battery drain mostly comes from data transfer and storage, not from sampling itself. GPS impacts battery much more than accelerometer data.

Harald
What’s the expected real-world accuracy?

Katharina
With 10,000 users over a year, we expect ~300 real accidents. We expect high accuracy for critical cases. Light parking damages may be missed.

Harald
You can adjust the ASI threshold based on your user base size and budget. Smaller groups can start at lower thresholds; larger ones need tighter controls due to call center costs.

Katharina
Yes. Literature says ASI around 1 corresponds to ~20% chance of moderate injury.

Harald
As an insurance company, you can fine-tune ASI based on road conditions and user behavior. For manual alarms, we use AI bots, but for crashes we recommend human review. Human eyes can spot context machines miss.

Katharina
Context is crucial.

Harald
To summarize: smartphones are not perfect, but they allow us to democratize accident detection. Second, we use a layered data science approach with both heuristics and machine learning. And third, we ensure early data capture with backend filtering and scalable platforms for security centers.

Katharina
Exactly.

Harald
Thank you, Katharina, for joining and helping us understand automatic impact detection using smartphone telematics.

Katharina
Thank you.