I am currently studying DSP and FFT, I am very new to this and have been doing electronics for a long time with Arduino and hobby projects. Recently, I am doing a project with the goal to map potholes along your daily commute in the car, measuring road roughness. This is not the road profile, but the subjected roughness of the ride the driver feels during the trip. I have an acceleration mounted to measure Z-axis "vertical" acceleration of the car, when a pothole is struck the shocks and springs dampen the force as per the quarter car model for simplicity.

Basically I want to be able to create a detector that will detect potholes by using FFT to find the dominating frequency of the vertical acceleration, then I can maybe match it to pattern to detect door slamming of the car door, or a man-hole cover from a pothole maybe in the future.

I'm not sure if FFT is the way to go, any input or advice and ideas are greatly appreciated, I have made a low pass filter to filter out the higher frequency vibration from the drive train and engine.

Any ideas on the subjected "roughness" felt by the passengers/driver in the car? I was thinking of doing road tests on a smooth road, gravel road, brick road, bad road, and coming up with a threshold of the magnitude of the vertical acceleration and making a scale somehow on what is considered "rough"

The potholes detected would be also logged with GPS their locations, I'm trying to use the Rasberry Pi for this, with data storage and its speed as a good project for it.

Thanks any help or ideas is appreciated, books, tutorials, wisdom, etc


Carry over additional data here:


  • 2
    $\begingroup$ PAK-9's answer below has some good suggestions. Step 1 for you should probably be to collect some data using road tests (as you suggested) to see what the features you would like to detect actually look like. $\endgroup$
    – Jason R
    Jan 4, 2013 at 17:34
  • $\begingroup$ Are you going to automatically submit them to seeclickfix.com? :) $\endgroup$
    – endolith
    Jan 4, 2013 at 18:24
  • $\begingroup$ Yes I could if I get it working correctly lol, The city of Boston made a smartphone app that does this, they had thousands of dollars in grants to people for their submissions. Mine is going to be a simple subset of that, I imagine with a smartphone most people move it around in their car, so you would have to have the phone in a holder to get accurate readings or detingish from dropping the phone or moving it while the car is moving. So I'm using an external accelerometer to avoid those complications $\endgroup$ Jan 4, 2013 at 18:30
  • $\begingroup$ @user978563 What is this external brand of accelerometer that you are using btw? $\endgroup$
    – Spacey
    Jan 4, 2013 at 18:57
  • $\begingroup$ Brand is KXPS5-3157 Datasheet dz863.com/downloadpdf-lovdnabsxg-KXPS5-3157.pdf $\endgroup$ Jan 4, 2013 at 19:09

1 Answer 1


You are not really interested in the frequency information in your signal so much as detectable features - as such I would suggest that there isn't that much utility in an FFT here.

What you really want to do is correlate the continuous 'car signal' recorded in the car with a different signal which is the 'signature' of the pothole bump. You will need to acquire this signature by recording it in a car going over a pothole (perhaps recording many and using them to produce a generic signal which is a good representation of a pothole). Once you have it you can cross correlate it with your car signal. You may want to do this in the frequency domain in which case you should cross correlate the FFT of your signature with a sliding FFT of the car signal.

You may want to process the car signal before you use it to remove, for example, high frequency noise (with a lowpass filter). You can do this pretty effectively by eye - play around with various filters on car signals to see which removes the most extraneous information while leaving the important pothole information intact.

  • $\begingroup$ Would this cross correlation be very hard to do programatically? I want this to be all done automatically in software. I guess the hardest part is to get the "generic signal" for a pothole broad enough to match the FFT against the FFT of the current data in the car. The sliding FFT, what would the windowing function be? The one thing that confuses me about FFT is the window, it would be a sliding window? It seems like the pothole frequencies occur in the 5HZ to 20HZ range very low frequency do to SHM of the dampened suspension $\endgroup$ Jan 4, 2013 at 17:21
  • 2
    $\begingroup$ Cross correlation is not especially difficult to do programatically, it is very similar to convolution (both of which are essentially iterating over two arrays and performing some function with a sample from each) - I am sure there are libraries available if you do not want to write your own. The signature may be a little difficult to acquire but if you get a load of data you should see some features emerging, plus cross correlation is not binary, the result is essentially a 'confidence' value so you can threshold it however you want. $\endgroup$
    – PAK-9
    Jan 4, 2013 at 17:29
  • 1
    $\begingroup$ The windowing fn of the sliding FFT can be whatever you want, it shouldn't affect the result too dramatically. I would expect to see a big impulse which is damped over a short period (SHM of the suspension as you say) so you may be able to synthesise something like this based on observations of the signals. $\endgroup$
    – PAK-9
    Jan 4, 2013 at 17:35
  • $\begingroup$ AH that makes sense, the difficult part is getting the signature is there any tutorial or book or site with a similar example of the procedure, right now I'm thinking I just need to make synthetic bumps kinda like speed bumps, and potholes and driver over them a bunch of times to get lots of data. I'm not sure what features I should consider for the signature though $\endgroup$ Jan 4, 2013 at 17:35
  • 2
    $\begingroup$ Just driving over a load of bumps and holes then chopping those parts of the signal out will give you a library of signatures. FFT each of them to give you a spectral signature, if some are widly different on visual inspection then seperate them out into groups which are similar. You can then perform some sort of averaging on each group of signatures to end up with a final 'master' signature from each group which you can compare against car signals. $\endgroup$
    – PAK-9
    Jan 4, 2013 at 17:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.