Today’s technology allows for fully automated vehicles to drive around everywhere and under all environmental conditions. If so, why is it that fully automated vehicles are not (yet) commercially sold?
One of the reasons is safety. Policies only allow for automated vehicles to drive on dedicated and protected roads: No humans allowed!
While this topic may cause intense debate, today I learnt that there is no obstacle detection method able to guarantee 100% detection accuracy. But wait a minute… Are human systems able to detect obstacles with 100% accuracy? Well if the answer was yes, we would never see road accidents…
Let’s talk about automated obstacle detection systems! In order for a computer to observe obstacles it requires an input source. Typical sources are cameras, lasers and sensors. Once the computer has its input, someone needs to program the computer to recognize obstacles in the input data.
Obstacles are all different but luckily they share similarities. The most typical similarities are shape properties, texture information and an innovative idea from Cheng-En Wu: self similarity.
In order to systematically compare all techniques, researchers selected specific obstacles according to their importance: pedestrians, cars or motorcycles. It was interesting to notice that no one chose bicycles (clearly they don’t live in the Netherlands).
The performance of each technique was measured based on positive obstacle detection and missed obstacles. The most impressive results in this conference were shown by Fernando García, who introduced a fusion methodology. He used both camera and laser input to detect vehicles and achieved 92% positive obstacle detection and 1% missed obstacles.
Challenges are still broad in the obstacle detection field. However, it is clear that methodologies are developing rapidly to enable fully automated vehicles to be commercialized in the near future. In the meantime, drivers can already benefit from the obstacle detection advice provided by their vehicles.