Hacker News Reader

Comment by iosjunkie | Hacker News Reader

Someone close to me was in an accident that could have been avoided with sensors other than a camera. The camera only collision mitigation system was blinded by direct sunlight complicated by glare off of fresh snow. Zero indication the system was compromised until it failed to provide any warning or active braking.

Vowed to never own a car with only camera based collision mitigation again.

View on HN

This. Boeing had removed a sensor deemed redundant for cost effectiveness, but failed to adapt the system to the changes and skipped the training of the pilots.

https://www.nytimes.com/2019/06/01/business/boeing-737-max-c...

How is this even remotely related to the topic? You just want to hate Boeing?

I think because it's a similar story that illustrates the danger of removing redundant sensors, like tesla is doing here

What was the driver doing?

That’s not really relevant, safety systems should be expected to work reliably, or report to the driver when they can’t (something my car does if covered in snow).

When radar and vision disagree, which one do you believe? Vision has much more precision, so better to double down on vision than do sensor fusion.

The solution two information from two sources disagreeing is rarely to just throw away half of your data.

In the event of a blinded vision system, such as sun glare, the vision system should be able to recognize its reduced capacity. That involves warning the driver and relying more heavily on radar/sensor fusion.

Like others have mentioned the entire field of sensor fusion deals with this problem. It is a very challenging problem to solve but it can be solved and it has been successfully used in spacecraft, aircraft, fighter jets, phones, AR/VR systems, and undoubtedly many others.

A basic approach is to have an uncertainty (or estimated uncertainty) for each of the sensing modalities. Then you use the uncertainties to weigh each sensing sample when deriving your estimated quantity (i.e. vehicle velocity for example). Assuming the uncertainties are correct, the resulting estimator can have variance lower than estimators deriving from a single sensor modality. Of course tuning sensor uncertainty values is a difficult problem in sensor fusion (and much more so when distributions are unknown) but it is definitely doable.

Repeating Elon's claim that sensor fusion is impossible/not-doable is entirely wrong. It is a technology that powers many different applications but its definitely not an easy thing to implement well.

There's a huge amount of literature on how to do sensor fusion in the presence of misbehaving or disagreeing sensors.

Moreover, production vehicles don't get entire classes of redundant sensors when the engineers think they can do without. The cost optimization and haggling that goes into production vehicles is insane.

Depends on your safety goal.

Reply on news.ycombinator.com