A colleague asked me last week if I thought it was possible to stick a couple GoPros on a pickup truck, drive through a collision site, and process the video in something like RealityCapture [RC] to generate a point cloud. I’d been wondering the same thing for a while, so I ran a few experiments this week.

Experiment #1: Two GoPro HERO7s were mounted to the upper corners of a Toyota Tundra’s windshield using Wood’s LJ45 suction cups (I’m a big fan). The truck was then piloted down the middle of a three-lane road, recording at 4K, 30 FPS.

The videos were imported into RC, extracting frames at various intervals (from every frame to every 6th), and aligned. Alignment was generally successful; however, the resultant point cloud was too sparse to be useful, no matter the number of frames used. A bit surprised, I moved to a new method.

Experiment #2: Thinking more camera separation would benefit the photogrammetric solution, one HERO7 was used for two passes on a two-lane road with a median. Eastbound: windshield mount. Westbound: tailgate mount, facing backward. This resulted in a lateral separation of about 24 feet between passes.

Again, the videos were imported into RC and frames were extracted at various intervals. Alignment was a little tricky, requiring several attempts and parametric adjustments, but it came together pretty well (see image below). Unfortunately, the resultant point cloud was overly sparse yet again.

Alignment of camera positions in RealityCapture.

Results were slightly improved with a HERO10, but not enough to call it a win. While success wasn’t achieved this week, I caught a glimpse and think it's possible.

I recently worked a case where the responding officer walked the path of a downed motorcycle, photographing evidence the whole way, and RC generated a beautiful point cloud of the scene. Granted, he was wielding a full-frame DSLR, but same concept. I think the combination of the lower quality imagery and moving vehicle in the experiments above prevented RC from identifying subtle, but distinct, patterns on the roadway, which are key informants for the solution.

If you’ve had success doing something similar, please let me know (I thought someone had accomplished this with DSLRs and published the results in SAE). Otherwise, I’ll probably use this as an excuse to grab the HERO13. Cranking that guy to 5.3K might do the trick. If not, photos or video from a DSLR or mirrorless camera would almost certainly get it done.

Thanks for reading, keep exploring!

Lou Peck
​Lightpoint | Axiom

P.S. Red means stop.