What's New in Axon's Crash Reconstruction Engine

For the better part of a decade, our crash reconstruction engine has been the quiet backbone of every FNOL alert, claims triage decision, and disputed-liability case that flows through an Axon-equipped fleet. This year, we rebuilt the core of it.

The new engine — rolling out across all newly deployed and updated devices — moves reconstruction from a rule-based, threshold-triggered process to a multi-sensor, event-segmented model. The result is faster, more granular, and, critically, more defensible in the claims and legal workflow.

This post is a technical deep dive into what changed and why it matters for insurers, claims teams, and fleet risk managers.

---

1. From thresholds to event segmentation

The previous generation of the engine worked the way most telematics crash detection still works today: a set of accelerometer and GPS thresholds, evaluated independently, each capable of triggering a "crash" event. The problem is well known to anyone who has worked in FNOL — threshold models are fast but brittle. A pothole, a kerb strike, a harsh brake, or a phone-mounted false reading can all look, to a single sensor in isolation, like an impact.

The new engine abandons single-threshold triggers in favour of event segmentation. Instead of asking "did any sensor cross a line?", it asks "what was the vehicle doing across a defined time window, and does the sensor signature across that window describe an impact?"

Concretely, every candidate event is now evaluated across three synchronized channels:

• ›High-rate IMU data (accelerometer + gyro, sampled at up to 100 Hz on supported hardware)
• ›GPS dynamics — speed, heading, and the second derivative of both, computed against the preceding baseline
• ›Vehicle bus signals — where available via CAN/OBD integration — including brake status, throttle position, steering angle, and seatbelt state

The engine stitches these into a single time-aligned event window, typically 12 seconds before and 6 seconds after the trigger candidate, then classifies the window as a whole.

---

2. Multi-axis impact classification

A reconstructed crash is not a single number. It is a sequence. The new engine breaks every event window into ordered phases and classifies each independently:

1. ›Pre-impact — the vehicle's state in the seconds before. Was it braking? Turning? Stationary? This anchors the "before" picture.
2. ›Primary impact — the peak delta-v event, identified by the largest sustained acceleration impulse along the dominant axis. The engine reports magnitude, direction, and duration.
3. ›Secondary events — subsequent impulses within the window. A rear-end push, a spin, a secondary collision. These are often the events that decide liability, and they are the ones threshold models most often miss.
4. ›Post-impact — the vehicle's trajectory after the event. Did it stop, roll, or continue? This distinguishes a low-speed shunt from a loss-of-control event.

Each phase carries its own confidence score, so a claims handler can see not just what happened but how sure we are.

---

3. Axis isolation and rollover detection

A frequent failure mode in single-axis telematics is the lateral impact that registers as a longitudinal event — or doesn't register at all. The new engine performs per-axis isolation, evaluating the acceleration vector in all three axes and resolving the dominant impact axis before classification.

This matters most for two scenarios that have historically been hard for black-box reconstruction:

• ›Side impacts at junctions, where the lateral impulse is large but the longitudinal component (what most legacy models key on) is small.
• ›Rollover events, where the gyro channel now carries the signal. The engine detects sustained rotation about the longitudinal axis and flags a rollover independent of any single linear acceleration peak.

For insurers, this means the scenarios that most often drive serious injury claims — side impacts and rollovers — are no longer the ones your telematics is least certain about.

---

4. Sensor fusion and the confidence model

No single sensor is trustworthy in isolation. GPS drifts in urban canyons. Accelerometers saturate. CAN data can drop out on older vehicles. The new engine runs a lightweight sensor-fusion layer that cross-validates each channel against the others and de-weights any channel that disagrees with the corroborating evidence.

The output is a composite confidence score for the whole event, expressed as a single value the claims team can act on:

• ›High confidence — multiple independent channels corroborate an impact. Triage as a confirmed crash.
• ›Medium confidence — primary channel confirms, secondary channels partial. Flag for review, but likely genuine.
• ›Low confidence — single-channel trigger with no corroboration. Suppress the FNOL alert and log for trend analysis rather than escalating.

This is the single biggest reason the new engine cuts false FNOL alerts so sharply: events that used to fire on one sensor now require agreement across the fusion layer before they ever reach a claims queue.

---

5. Timing and delta-v accuracy

For reconstruction that needs to hold up in a disputed claim or a legal setting, the two questions that matter most are when and how hard.

Timing. The engine now timestamps every phase to the millisecond against device UTC, with a disciplined clock-discipline loop that keeps the device within tight drift bounds even through long connectivity gaps. When a crash report says the primary impact occurred at 14:07:32.418, that timestamp is forensic-grade, not approximate.

Delta-v. Peak delta-v — the change in velocity through the impact — is now integrated across the full impact phase rather than sampled at the peak instant. This gives a defensible, repeatable measure of crash severity that correlates far better with injury risk and repair cost than the single-instant figures produced by threshold models.

---

6. What this means in practice

For an insurer running Axon-equipped policies, the new engine delivers three measurable shifts:

• ›Fewer false FNOL alerts. The fusion-and-confidence model suppresses the single-sensor triggers that account for the bulk of wasted claims-team time.
• ›Better liability calls. Multi-phase, multi-axis reconstruction gives the before-during-after picture a claims handler needs to settle a disputed case without a site visit.
• ›Defensible severity. Integrated delta-v and millisecond timing produce evidence that stands up in court, not just on a dashboard.

For fleet operators, the same engine feeds directly into driver-coaching workflows: secondary events and near-misses that were previously lost in the noise now surface as actionable, coachable incidents.

---

Rollout

The new reconstruction engine is live on all newly deployed Axon devices and is being pushed to the installed base through the standard firmware update channel. No hardware change is required for the majority of supported devices — the engine runs on existing IMU and GPS capability, and simply unlocks additional precision where CAN/OBD signals are available.

If you'd like a walkthrough of the new engine against your own historical event data, or a technical briefing for your claims and actuarial teams, contact us — we're happy to show you exactly what changed, against the events you already know.