The one thing nobody tells you about dashcam loop recording
Three weeks after launch, I got an email from a driver in Manchester. He'd been using Hawk for a month, had crisp footage of a junction collision, and was ready to send it to his insurer. Then he watched it back. The video wobbled. Not catastrophically, but enough that he worried it would undermine his claim. That message stuck with me because it exposed something the entire dashcam space glosses over: continuous recording is pointless if your footage looks like you filmed it while driving over cobblestones.
The problem with filming from a moving vehicle
When your phone is mounted on a dashboard, it's experiencing constant micro-vibrations. Engine vibration, road texture, suspension movement. Over a 10-minute continuous recording, those vibrations compound. Normal video stabilisation works by cropping and shifting the frame slightly, but it needs to know where to shift. It looks for visual features in consecutive frames and tracks them. On a highway, with scenery flowing past and trees bouncing, that's feasible. But in traffic, especially at low speeds with repetitive shapes (other cars, lane markings, streetlights), traditional stabilisation struggles. It either over-corrects, creating a nauseating drift, or under-corrects, leaving the wobble.
The Manchester driver's footage had that wobble. Not his fault. Not the phone's fault. Just the physics of mounting a sensor on something that never stops moving.
What optical-flow stabilisation actually does differently
Optical flow is a computer-vision technique that doesn't just track point-like features. It models the motion of every pixel across the frame, building a dense map of which parts of the image moved and by how much. Think of it as understanding not just where a car is, but how the light and shadow across the entire road surface is shifting. That granularity matters when you're trying to separate the true motion of the road (which you want to keep) from the vibration of the phone (which you want to remove).
The stabilisation algorithm uses that flow map to predict what the next frame should look like if only the camera had moved smoothly, then warps the actual frame to match that prediction. It's not reframing the video. It's applying a subtle warp field that's calculated per frame. The result looks like the footage was shot from a steady hand, not from a vibrating dashboard.
For evidence purposes, this matters. Police and insurers expect to see clear footage. A wobbly clip looks amateur, even if the facts in it are rock-solid. A stable clip looks professional and credible. That's not superficial. Courts and claims handlers make snap judgements based on presentation.
Why we chose this over cropping or just accepting wobble
When we were building Hawk, we could have solved the vibration problem the simple way. Crop the frame, apply a basic motion-tracking stabiliser, call it done. It would have been faster. It would have been cheaper. We'd have shipped three months earlier.
But we kept coming back to the same question: if someone uses their phone as a dashcam, they're usually not doing it because they want a fancy device. They're doing it because they already have a phone in their pocket. The only thing they're sacrificing is footage quality. It felt wrong to then give them footage that looked like it was filmed in an earthquake. So we went deeper.
Optical-flow stabilisation is computationally heavier. It runs during recording on the device itself, not in post. That meant we had to optimise it ruthlessly to not drain the battery or cause thermal issues during a six hour rideshare shift. It meant extensive testing across iPhone models, from older A-series chips to the latest. And yes, it meant delaying launch by about six weeks.
The Manchester driver watched his re-stabilised footage and sent it to his insurer the same day. Claim approved within a week. That's the difference a clear picture makes.
How stabilisation lives inside evidence-grade recording
Here's something people don't always realize: stabilisation usually happens after recording, in software, which means the original file is never touched. That's fine for holiday videos. It's not fine for evidence. If the footage has been processed, there's a chain-of-custody question. Did the stabilisation introduce artifacts? Could it have been manipulated further? Courts and police have started asking these questions.
In Hawk, the optical-flow stabilisation runs in real-time, during capture, but the frame data is recorded before warping is applied. Every clip is written with a SHA-256 integrity hash that covers the stabilised video. That hash is cryptographic proof that the file hasn't changed since it was recorded. When you export a dispute to your insurer or police, that manifest and those hashes come with it. You're not just handing over stabilised footage. You're handing over proof that the footage is authentic and unaltered since the moment it was recorded.
It's the difference between saying, 'Here's a video that looks good', and saying, 'Here's a video that hasn't been touched since the dashcam locked it'.
The ripple effect: why this matters for long drives
Continuous loop recording means you're not managing clips. The phone records, stores up to 10 minutes per file (by default), and when storage fills up, it starts overwriting the oldest footage. That's where stabilisation compounds its value. On a 90-minute drive, you could have nine separate 10-minute clips. Each one stabilised in real-time. No gaps. No stitching artifacts. Just one unbroken, credible record of the journey.
A rideshare driver working a night shift has six or seven trips stacked in their log. Each one its own clip, each one with its own GPS trace and timestamp. Each one stable enough to stand up to scrutiny. If a passenger claims you hit a pothole or drove recklessly, you have proof. Not a jumpy, unconvincing proof. Real proof.
That's continuous recording done right. Not just recording everything. Recording everything in a way that nobody can question.
The next time you watch a dashcam video online, pay attention to the wobble. Then ask yourself whether you'd trust that footage in a dispute. That's what we're solving.