Hacker News Reader

Comment by onionisafruit | Hacker News Reader

How did you fix that? I’ve been wondering how google and the others differentiate between cars on the free vs express lanes.

View on HN

You can measure peak acceleration--cars accelerate much faster than buses or trains--or whether you're connected to headphones (not a head-unit). Use of headphones by drivers is illegal in most places.

As for express vs. regular lanes, express-lane cars will have much different distributions than regular-lane cars. Use something like KNN where the distance metric is a weighted sum of Kolmogorov-Smirnov distance between speed distributions and K-S between acceleration distributions (each vehicle is one unit with a distribution of speeds and accelerations).

> or whether you're connected to headphones

I’m pretty sure this is rarely enforced in most states in the US. My father frequently drives across the US and always takes his calls with earbuds because the speakers aren’t very good.

I do it with one earbud in because I find the quality of speaker audio terrible, forcing me to concentrate harder on it to discern what’s being said.

Single ear devices are usually legal where earbuds in both ears is not. Maintaining situational awareness seems to be the rule. If course, I'm essentially deaf in on ear, so being legal isn't any safer.

I tend to use the car speakers or simply refuse all incoming calls while driving.

> simply refuse all incoming calls while driving

Definitely the best course of action. Even just talking on the phone is very distracting. No way someone can have a full conversation going and still be giving the same attention to driving.

Or just split speeds into a histogram and take the 20th percentile or something.

I don't know what a free lane vs an express lane is, but for the "that", I can imagine that a cluster of hundreds of phones moving at exactly the same speed in the vicinity of a train track until they reach the next station would be a reasonable indicator that the were in a train. The preexisting phones in that area would not show the property of moving at the same speed at the same time, and it also kind of stands out that they're still doing 15-30km/h (or however bad traffic may be) while the newcomer phones all rush through at 140km/h (85miles/h).

Bonus points if those clients all connect from the same IP address (range). That would be the train WiFi.

Express lanes can be separated lanes with fewer exits that allow you to go faster because of fewer lane changes. Chicago (and I'm sure many other cities) have these. They're separated by concrete barriers and only have exits every couple miles, so you'll be right next to super slow traffic while you're doing 20+ MPH more than the "free" lanes. In Chicago, you don't have to pay to go in them, but here in LA, they have High Occupancy Vehicle lanes that you can use if you have more than one person in the, OR if you pay for using the lane with the FastPass you use for paying for other tolls.

You could collect all data from that route and then if you find that the average velocities of the phones fall neatly into two sets of values, then the algorithm should interpret this as one set of phones having an alternate method of travel, something other than the road.

That's quite easy and there are different ways to do it, the first that comes to mind is using location history.

I used to work for a map company, and phone traces were immediately broken into smaller chunks to anonymize them. Using location history isn’t viable when all you have a collection of short traces.

In audio processing you deal with spikes using one or two averaging filters with different time constants for rise/fall.

You can also do it with a PID approach

The device itself should be able to discern the mode of transportation, no?

Indeed, but carriers don't get data from the device's sensors, so they need to be a little smart to properly recognize the mode of transportation.

It's easier with historical data since you know the full trip, but even real-time should be doable.

This should be simple if there’s a beacon or Wifi on the train. Train goers will have that beacon in range on each location update while car goers will not (or may briefly if in-range of the train)

Pretty easy to train some ML network to differentiate between transport modes, assuming the device has a accelerometer / gyroscope / whatever.

Not really that easy with just IMU sensors, I’d assume ?

I did one project in school (albeit a simple SVM based model) to classify walking vs running (and so on) and ambiguity was still there. Stationary vs. Walking was easy to draw a hyper plane in between, not Running.

I still find it a difficult problem to get into, after 5 years. Perhaps ANN models might work better? Although my heart is still at Hidden MM states...

Do you have the code for your classification project? It would be great to see it on github or something.

I have it, not only the code but labelled dataset I created with an arduino, first, and then iPhone. It’s been a private repository forever, but thanks! I’ll make it public and Edit post the link down here.

(Its not pretty though...)

Edit: https://github.com/prashnts/MPU-9250

I have proper project reports I’d submitted somewhere for sure, as well. I’ll add it there in repo. I never bothered earlier.

Edit 2: this was the paper i finally submitted at school.

https://github.com/prashnts/MPU-9250/blob/master/Docs/submit...

I enjoyed reading that, thank you!

Thanks! To be more honest, this paper had a "constraint" against using FFT, simply because I didn't really know how it worked then...

Well if the people are actually navigating, in GMaps they normally have to select which mode they're using: car, bicycle, public trans, walking, etc.

Pick the 10%ile speed, assume that people sent doing that voluntarily.

Reply on news.ycombinator.com