A camera-only self-driving car is like ChatGPT on wheels – dazzling when it works well, but not immune to hallucination. For the trucking industry, the camera-only concept is a fantasy that ends in a crash course in physics. I’ll explain why.
For any autonomous driving system, knowing how far away objects are and their relative velocity is critical. A car doing 70 mph needs roughly 100 metres to stop in an emergency, and out to that distance, camera-based depth estimates are generally good enough. (The keyword there is “estimates”, because they are not measurements, but educated guesses.)
Camera systems estimate depth by triangulation. That means – as anyone with two eyes can tell you – the further away something is, the harder it is to precisely gauge its distance or speed. It’s an unavoidable mathematical reality that with triangulation, uncertainty doesn’t grow linearly, but quadratically. In other words, double the distance and you quadruple the error.
At a range of 200 metres – the distance a fully-loaded Class 8 truck travelling at highway speeds needs to stop – you simply can’t trust cameras any more. For example, a camera-based system might be accurate to within 8 meters at a range of 100m, but off by more than 30m at a range of 200m. That’s the difference between stopping safely and ploughing through an obstacle whose position and speed you misjudged. (....)