A team of researchers has built an automated tool that blurs faces and license plates in Kuala Lumpur's chaotic, motorbike-heavy street footage well enough to make the video usable for training self-driving systems.
The researchers were curating the Kuala Lumpur Road Dataset, filmed from a mobile cycling platform at 2 frames per second, and ran into a specific mess: dense motorcycle traffic, dark acrylic license plates, a camera that tilts as it rides, and glare from tropical sun. Older detection methods, including Haar cascades and YOLOv8, choked on these conditions, flagging background clutter as license plates while missing ones that were rotated or partly hidden. The new pipeline pairs Grounding DINO, a zero-shot open-set vision-language detector, with what the team calls a Spatial Vehicle Region of Interest Containment Engine, which only accepts a license plate as real if its center falls inside a detected vehicle's boundary. Across 1,266 test frames, the system anonymized faces, heads, and plates with about a 95 percent success rate, with most remaining misses limited to small, heavily occluded, or oblique targets.
This matters because most privacy-anonymization tools were built and tuned on tidy Western dashcam footage: fewer motorcycles, brighter plates, predictable camera angles. Southeast Asian cities look nothing like that, and a tool that only works on orderly traffic can't ethically anonymize the messier, denser streets where a lot of the world actually drives. A pipeline built for Kuala Lumpur's specific chaos is a template other tropical or high-density cities could reuse.
Ninety-five percent success across 1,266 frames is a solid proof of concept, not a finished product. The researchers say as much themselves: passing this internal benchmark is not the same as passing a real privacy audit, since legal compliance still depends on governance work they haven't done yet.