Lab notes · Road safety
Seeing cones is not the same as understanding a lane closure
Robotaxis have been recalled after driving between cones into active freeway work zones. Not because they couldn't see the cones. Detecting an orange cone is a solved task. The failure is subtler: seeing cones is not the same as understanding a lane closure.
A cone is an object. A lane closure is a decision: which lane is shut, which way traffic has to move, and how much distance there is to do it. Nothing about a bounding box tells you that.
So we built the missing layer. Dashcam footage goes in. What comes out is the merge line drawn through the cones, the closed lane shaded in perspective, the merge direction, and the distance to the taper in metres.
Two details we like. First, no training data was collected for this: the detector is simply prompted with the words “traffic cone”. The test clip was black and white, so there was no orange color to lean on — it worked from shape and geometry alone. Second, the system checks itself: from the cone sizes alone it recovers the camera's mounting height, and lands on a realistic dashcam height that nobody gave it.
The gap between “I see cones” and “I understand this closure” is where road-safety value lives.
Capability demo built by IPD Scientific. More lab notes →
Have footage like this?
Book a free 15-minute call, or email us a sample image or short clip. We'll tell you honestly whether computer vision can do the job.