Related
I am new to Android and not familiar with the different type of sensors. I am working on an App and part of it has to count the number of speedbreakers that a car will pass over during its journey.
The phone will remain stationary in one position
I have tried using the accelerometer and tried to use a peak in vertical acceleration followed by a negative vertical acceleration as an indicator of a speedbreaker but there is too much fluctuation for an accurate result.
Here speedbreakers are smooth slopes of cement, usually a few inches high.
Any help or guidance would be greatly appreciated.
I would start out with logging the accelerometer data (probably the speed data (from gps) as well) and manually marking the points where you pass a speedbreaker.
Then the first step would be to see if there is something to see here: maybe there is a clear 'signal' that's apart from the normal fluctuation, but you haven't gotten the tweaking right.
If there isn't, you can always see if there is something there that you haven't recognized. Some sort of of normal behaviour that stops for a bit. These can be harder to detect visually so you'd have to do something with the signal.
If you know nothing about signal processing it might be tricky, but as a random starting point, read up on how step-detection works: https://en.wikipedia.org/wiki/Step_detection
Some of the methods might be usefull to you. Look at the FFT, process your signal to filter out the points you need. Maybe even train a simple network to see if it finds anything going on at your desired points?
I was looking into implementing an Inertial Navigation System for an Android phone, which I realise is hard given the accelerometer accuracy, and constant fluctuation of readings.
To start with, I set the phone on a flat surface and sampled 1000 accelerometer readings in the X and Y directions (parallel to the table, so no gravity acting in these directions). I then averaged these readings and used this value to calibrate the phone (subtracting this value from each subsequent reading).
I then tested the system by again placing it on the table and sampling 5000 accelerometer readings in the X and Y directions. I would expect, given the calibration, that these accelerations should add up to 0 (roughly) in each direction. However, this is not the case, and the total acceleration over 5000 iterations is nowhere near 0 (averaging around 10 on each axis).
I realise without seeing my code this might be difficult to answer but in a more general sense...
Is this simply an example of how inaccurate the accelerometer readings are on a mobile phone (HTC Desire S), or is it more likely that I've made some errors in my coding?
You get position by integrating the linear acceleration twice but the error is horrible. It is useless in practice.
Here is an explanation why (Google Tech Talk) at 23:20. I highly recommend this video.
It is not the accelerometer noise that causes the problem but the gyro white noise, see subsection 6.2.3 Propagation of Errors. (By the way, you will need the gyroscopes too.)
As for indoor positioning, I have found these useful:
RSSI-Based Indoor Localization and Tracking Using Sigma-Point Kalman Smoothers
Pedestrian Tracking with Shoe-Mounted Inertial Sensors
Enhancing the Performance of Pedometers Using a Single Accelerometer
I have no idea how these methods would perform in real-life applications or how to turn them into a nice Android app.
A similar question is this.
UPDATE:
Apparently there is a newer version than the above Oliver J. Woodman, "An introduction to inertial navigation", his PhD thesis:
Pedestrian Localisation for Indoor Environments
I am just thinking out loud, and I haven't played with an android accelerometer API yet, so bear with me.
First of all, traditionally, to get navigation from accelerometers you would need a 6-axis accelerometer. You need accelerations in X, Y, and Z, but also rotations Xr, Yr, and Zr. Without the rotation data, you don't have enough data to establish a vector unless you assume the device never changes it's attitude, which would be pretty limiting. No one reads the TOS anyway.
Oh, and you know that INS drifts with the rotation of the earth, right? So there's that too. One hour later and you're mysteriously climbing on a 15° slope into space. That's assuming you had an INS capable of maintaining location that long, which a phone can't do yet.
A better way to utilize accelerometers -even with a 3-axis accelerometer- for navigation would be to tie into GPS to calibrate the INS whenever possible. Where GPS falls short, INS compliments nicely. GPS can suddenly shoot you off 3 blocks away because you got too close to a tree. INS isn't great, but at least it knows you weren't hit by a meteor.
What you could do is log the phones accelerometer data, and a lot of it. Like weeks worth. Compare it with good (I mean really good) GPS data and use datamining to establish correlation of trends between accelerometer data and known GPS data. (Pro tip: You'll want to check the GPS almanac for days with good geometry and a lot of satellites. Some days you may only have 4 satellites and that's not enough) What you might be able to do is find that when a person is walking with their phone in their pocket, the accelerometer data logs a very specific pattern. Based on the datamining, you establish a profile for that device, with that user, and what sort of velocity that pattern represents when it had GPS data to go along with it. You should be able to detect turns, climbing stairs, sitting down (calibration to 0 velocity time!) and various other tasks. How the phone is being held would need to be treated as separate data inputs entirely. I smell a neural network being used to do the data mining. Something blind to what the inputs mean, in other words. The algorithm would only look for trends in the patterns, and not really paying attention to the actual measurements of the INS. All it would know is historically, when this pattern occurs, the device is traveling and 2.72 m/s X, 0.17m/s Y, 0.01m/s Z, so the device must be doing that now. And it would move the piece forward accordingly. It's important that it's completely blind, because just putting a phone in your pocket might be oriented in one of 4 different orientations, and 8 if you switch pockets. And there's many ways to hold your phone, as well. We're talking a lot of data here.
You'll obviously still have a lot of drift, but I think you'd have better luck this way because the device will know when you stopped walking, and the positional drift will not be a perpetuating. It knows that you're standing still based on historical data. Traditional INS systems don't have this feature. The drift perpetuates to all future measurements and compounds exponentially. Ungodly accuracy, or having a secondary navigation to check with at regular intervals, is absolutely vital with traditional INS.
Each device, and each person would have to have their own profile. It's a lot of data and a lot of calculations. Everyone walks different speeds, with different steps, and puts their phones in different pockets, etc. Surely to implement this in the real world would require number-crunching to be handled server-side.
If you did use GPS for the initial baseline, part of the problem there is GPS tends to have it's own migrations over time, but they are non-perpetuating errors. Sit a receiver in one location and log the data. If there's no WAAS corrections, you can easily get location fixes drifting in random directions 100 feet around you. With WAAS, maybe down to 6 feet. You might actually have better luck with a sub-meter RTK system on a backpack to at least get the ANN's algorithm down.
You will still have angular drift with the INS using my method. This is a problem. But, if you went so far to build an ANN to pour over weeks worth of GPS and INS data among n users, and actually got it working to this point, you obviously don't mind big data so far. Keep going down that path and use more data to help resolve the angular drift: People are creatures of habit. We pretty much do the same things like walk on sidewalks, through doors, up stairs, and don't do crazy things like walk across freeways, through walls, or off balconies.
So let's say you are taking a page from Big Brother and start storing data on where people are going. You can start mapping where people would be expected to walk. It's a pretty sure bet that if the user starts walking up stairs, she's at the same base of stairs that the person before her walked up. After 1000 iterations and some least-squares adjustments, your database pretty much knows where those stairs are with great accuracy. Now you can correct angular drift and location as the person starts walking. When she hits those stairs, or turns down that hall, or travels down a sidewalk, any drift can be corrected. Your database would contain sectors that are weighted by the likelihood that a person would walk there, or that this user has walked there in the past. Spatial databases are optimized for this using divide and conquer to only allocate sectors that are meaningful. It would be sort of like those MIT projects where the laser-equipped robot starts off with a black image, and paints the maze in memory by taking every turn, illuminating where all the walls are.
Areas of high traffic would get higher weights, and areas where no one has ever been get 0 weight. Higher traffic areas are have higher resolution. You would essentially end up with a map of everywhere anyone has been and use it as a prediction model.
I wouldn't be surprised if you could determine what seat a person took in a theater using this method. Given enough users going to the theater, and enough resolution, you would have data mapping each row of the theater, and how wide each row is. The more people visit a location, the higher fidelity with which you could predict that that person is located.
Also, I highly recommend you get a (free) subscription to GPS World magazine if you're interested in the current research into this sort of stuff. Every month I geek out with it.
I'm not sure how great your offset is, because you forgot to include units. ("Around 10 on each axis" doesn't say much. :P) That said, it's still likely due to inaccuracy in the hardware.
The accelerometer is fine for things like determining the phone's orientation relative to gravity, or detecting gestures (shaking or bumping the phone, etc.)
However, trying to do dead reckoning using the accelerometer is going to subject you to a lot of compound error. The accelerometer would need to be insanely accurate otherwise, and this isn't a common use case, so I doubt hardware manufacturers are optimizing for it.
Android accelerometer is digital, it samples acceleration using the same number of "buckets", lets say there are 256 buckets and the accelerometer is capable of sensing from -2g to +2g. This means that your output would be quantized in terms of these "buckets" and would be jumping around some set of values.
To calibrate an android accelerometer, you need to sample a lot more than 1000 points and find the "mode" around which the accelerometer is fluctuating. Then find the number of digital points by how much the output fluctuates and use that for your filtering.
I recommend Kalman filtering once you get the mode and +/- fluctuation.
I realise this is quite old, but the issue at hand is not addressed in ANY of the answers given.
What you are seeing is the linear acceleration of the device including the effect of gravity. If you lay the phone on a flat surface the sensor will report the acceleration due to gravity which is approximately 9.80665 m/s2, hence giving the 10 you are seeing. The sensors are inaccurate, but they are not THAT inaccurate! See here for some useful links and information about the sensor you may be after.
You are making the assumption that the accelerometer readings in the X and Y directions, which in this case is entirely hardware noise, would form a normal distribution around your average. Apparently that is not the case.
One thing you can try is to plot these values on a graph and see whether any pattern emerges. If not then the noise is statistically random and cannot be calibrated against--at least for your particular phone hardware.
I'm trying to write an android code that uses the device sensors to detect freefall scenario.
I searched the web a lot for a solution to this problem but I was unable to find anything useful.
I did see that there are several apps that does exactly this so it is possible, but i didn't find any code sample or tutorial on how to do that.
Can anyone please help me with a code snippet or even with a mathematical calculation using the sensors data?
Thanks in advance
The device is in free fall if the length of the vector given by TYPE_ACCELEROMETER is approximately zero. In theory, it should be exactly zero, in practice, it well be only near zero. So you need to come up with some threshold by trial and error and declare that the device is in free fall if the length of that vector is below this threshold.
Check out the API here SensorEvent Values and the math behind FreeFall here Wikipedia.
You are trying to detect speed in some direction. Look at the motion equations on wikipedia. You are detecting acceleration over time that is normalized for the gyroscopic rotation of the device.
Also see: How to approach Fall Detection Algorithm
From what I've read, the accelerometer normally measures gravity. Thus, if you're in freefall and the device is not being moved laterally, all accelerometer readings should be zero. (Disclaimer: I have not written any accelerometer code.)
Google the iFall project by a group at Florida State University. The have published a paper describiing the approach they took for their Android iFall application, which gives a host of references for further/extended study. They also have an API available and explain how to use it, if you want a fast shortcut approach. (To use their API, I believe you just need to download and install their iFall app form the Playstore)
I'm building an application for Android devices that requires it to recognize, by accelerometer data, the difference between walking noise and double tapping it. I'm trying to solve this problem using Neural Networks.
At the start it went pretty well, teaching it to recognize the taps from noise such as standing up/ sitting down and walking around at a slower pace. But when it came to normal walking it never seemed to learn even though I fed it with a large proportion of noise data.
My question: Are there any serious flaws in my approach? Is the problem based on lack of data?
The network
I've choosen a 25 input 1 output multi-layer perceptron, which I am training with backpropagation. The input is the changes in acceleration every 20ms and output ranges from -1 (for no-tap) to 1 (for tap). I've tried pretty much every constallation of hidden inputs there are, but had most luck with 3 - 10.
I'm using Neuroph's easyNeurons for the training and exporting to Java.
The data
My total training data is about 50 pieces double taps and about 3k noise. But I've also tried to train it with proportional amounts of noise to double taps.
The data looks like this (ranges from +10 to -10):
Sitting double taps:
Fast walking:
So to reiterate my questions: Are there any serious flaws in my approach here? Do I need more data for it to recognize the difference between walking and double tapping? Any other tips?
Update
Ok so after much adjusting we've boiled the essential problem down to being able to recognize double taps while taking a brisk walk. Sitting and regular (in-house) walking we can solve pretty good.
Brisk walk
So this is some test data of me first walking then stopping, standing still, then walking and doing 5 double taps while I'm walking.
If anyone is interested in the raw data, I linked it for the latest (brisk walk) data here
Do you insist on using a neural network? If not, here is an idea:
Take a window of 0.5 seconds and consider the area under the curve (or since your signal is discrete, the sum of the absolute values of each sensor reading-- the red area in the attached image). You will probably find that that sum is high when the user is walking and much much lower when they are sitting and/or tapping. You can set a threshold above which you consider a given window to be taken while the user is walking. Alternatively, since you have labelled data, you can train any binary classifier to differentiate between walking and not walking.
You can probably improve your system by considering other features of the signal, such as how jagged the line is. If the phone is sitting on a table, the line will be almost flat. If the user is typing, the line will be kind of flat, and you will see a spike every now and then. If they are walking, you will see something like a sine wave.
Have you considered that the "fast walking" and "fast walking + double tapping" signals might be too similar to differentiate using only accelerometer data? It may simply not be possible to achieve accuracy above a certain amount.
Otherwise, neural networks are probably a good choice for your data, and it still may be possible to get better performance out of them.
This very-useful paper (http://yann.lecun.com/exdb/publis/pdf/lecun-98b.pdf) recommends that you whiten your dataset so that it has a mean of zero and unit covariance.
Also, since your problem is a classification problem, you should make sure that you are training your network using a cross-entropy criteria (http://arxiv.org/pdf/1103.0398v1.pdf ) rather than RMSE. (I have no idea whether Neuroph supports cross-entropy or not.)
Another relatively simple thing you could try, as other posters suggested, is transforming your data. Using an FFT or DCT to transform your data to the frequency domain is relatively standard for time-series classification.
You could also try training networks on different sized windows and averaging the results.
If you want to try some more difficult NN architectures, you could look at the Time-Delay-Neural-Network (just google this for the paper), which takes multiple windows into account in its structure. It should be relatively straightforward to use one of the Torch libraries (http://www.torch.ch/) to implement this, but it might be hard to export the network to an Android environment.
Finally, another method of getting better classification performance in time-series data is to consider the relationships between adjacent labels. Conditional Neural Fields (http://code.google.com/p/cnf/ - note:I have never used this code) do this by integrating neural networks into conditional random fields, and, depending on the patterns of behavior in your actual data, may do a better job.
What probably would work is to filter the data using a Fourier transform first. Walking has a sinus like amplitude, your double taps would stand-out in the transform-result as a different frequency. I guess a neural network can than determine if the data contains your double tabs because it has the extra frequency (the double tabs frequency). Some questions remain:
How long the sample of data needs to be?
Can your phone do all the work it needs to do, does it have enough processing power?
You might even want to consider using the GPU for this.
Another option is to use the Fourier output and some good old Fuzzy Logic.
This sound like fun...
I was looking into implementing an Inertial Navigation System for an Android phone, which I realise is hard given the accelerometer accuracy, and constant fluctuation of readings.
To start with, I set the phone on a flat surface and sampled 1000 accelerometer readings in the X and Y directions (parallel to the table, so no gravity acting in these directions). I then averaged these readings and used this value to calibrate the phone (subtracting this value from each subsequent reading).
I then tested the system by again placing it on the table and sampling 5000 accelerometer readings in the X and Y directions. I would expect, given the calibration, that these accelerations should add up to 0 (roughly) in each direction. However, this is not the case, and the total acceleration over 5000 iterations is nowhere near 0 (averaging around 10 on each axis).
I realise without seeing my code this might be difficult to answer but in a more general sense...
Is this simply an example of how inaccurate the accelerometer readings are on a mobile phone (HTC Desire S), or is it more likely that I've made some errors in my coding?
You get position by integrating the linear acceleration twice but the error is horrible. It is useless in practice.
Here is an explanation why (Google Tech Talk) at 23:20. I highly recommend this video.
It is not the accelerometer noise that causes the problem but the gyro white noise, see subsection 6.2.3 Propagation of Errors. (By the way, you will need the gyroscopes too.)
As for indoor positioning, I have found these useful:
RSSI-Based Indoor Localization and Tracking Using Sigma-Point Kalman Smoothers
Pedestrian Tracking with Shoe-Mounted Inertial Sensors
Enhancing the Performance of Pedometers Using a Single Accelerometer
I have no idea how these methods would perform in real-life applications or how to turn them into a nice Android app.
A similar question is this.
UPDATE:
Apparently there is a newer version than the above Oliver J. Woodman, "An introduction to inertial navigation", his PhD thesis:
Pedestrian Localisation for Indoor Environments
I am just thinking out loud, and I haven't played with an android accelerometer API yet, so bear with me.
First of all, traditionally, to get navigation from accelerometers you would need a 6-axis accelerometer. You need accelerations in X, Y, and Z, but also rotations Xr, Yr, and Zr. Without the rotation data, you don't have enough data to establish a vector unless you assume the device never changes it's attitude, which would be pretty limiting. No one reads the TOS anyway.
Oh, and you know that INS drifts with the rotation of the earth, right? So there's that too. One hour later and you're mysteriously climbing on a 15° slope into space. That's assuming you had an INS capable of maintaining location that long, which a phone can't do yet.
A better way to utilize accelerometers -even with a 3-axis accelerometer- for navigation would be to tie into GPS to calibrate the INS whenever possible. Where GPS falls short, INS compliments nicely. GPS can suddenly shoot you off 3 blocks away because you got too close to a tree. INS isn't great, but at least it knows you weren't hit by a meteor.
What you could do is log the phones accelerometer data, and a lot of it. Like weeks worth. Compare it with good (I mean really good) GPS data and use datamining to establish correlation of trends between accelerometer data and known GPS data. (Pro tip: You'll want to check the GPS almanac for days with good geometry and a lot of satellites. Some days you may only have 4 satellites and that's not enough) What you might be able to do is find that when a person is walking with their phone in their pocket, the accelerometer data logs a very specific pattern. Based on the datamining, you establish a profile for that device, with that user, and what sort of velocity that pattern represents when it had GPS data to go along with it. You should be able to detect turns, climbing stairs, sitting down (calibration to 0 velocity time!) and various other tasks. How the phone is being held would need to be treated as separate data inputs entirely. I smell a neural network being used to do the data mining. Something blind to what the inputs mean, in other words. The algorithm would only look for trends in the patterns, and not really paying attention to the actual measurements of the INS. All it would know is historically, when this pattern occurs, the device is traveling and 2.72 m/s X, 0.17m/s Y, 0.01m/s Z, so the device must be doing that now. And it would move the piece forward accordingly. It's important that it's completely blind, because just putting a phone in your pocket might be oriented in one of 4 different orientations, and 8 if you switch pockets. And there's many ways to hold your phone, as well. We're talking a lot of data here.
You'll obviously still have a lot of drift, but I think you'd have better luck this way because the device will know when you stopped walking, and the positional drift will not be a perpetuating. It knows that you're standing still based on historical data. Traditional INS systems don't have this feature. The drift perpetuates to all future measurements and compounds exponentially. Ungodly accuracy, or having a secondary navigation to check with at regular intervals, is absolutely vital with traditional INS.
Each device, and each person would have to have their own profile. It's a lot of data and a lot of calculations. Everyone walks different speeds, with different steps, and puts their phones in different pockets, etc. Surely to implement this in the real world would require number-crunching to be handled server-side.
If you did use GPS for the initial baseline, part of the problem there is GPS tends to have it's own migrations over time, but they are non-perpetuating errors. Sit a receiver in one location and log the data. If there's no WAAS corrections, you can easily get location fixes drifting in random directions 100 feet around you. With WAAS, maybe down to 6 feet. You might actually have better luck with a sub-meter RTK system on a backpack to at least get the ANN's algorithm down.
You will still have angular drift with the INS using my method. This is a problem. But, if you went so far to build an ANN to pour over weeks worth of GPS and INS data among n users, and actually got it working to this point, you obviously don't mind big data so far. Keep going down that path and use more data to help resolve the angular drift: People are creatures of habit. We pretty much do the same things like walk on sidewalks, through doors, up stairs, and don't do crazy things like walk across freeways, through walls, or off balconies.
So let's say you are taking a page from Big Brother and start storing data on where people are going. You can start mapping where people would be expected to walk. It's a pretty sure bet that if the user starts walking up stairs, she's at the same base of stairs that the person before her walked up. After 1000 iterations and some least-squares adjustments, your database pretty much knows where those stairs are with great accuracy. Now you can correct angular drift and location as the person starts walking. When she hits those stairs, or turns down that hall, or travels down a sidewalk, any drift can be corrected. Your database would contain sectors that are weighted by the likelihood that a person would walk there, or that this user has walked there in the past. Spatial databases are optimized for this using divide and conquer to only allocate sectors that are meaningful. It would be sort of like those MIT projects where the laser-equipped robot starts off with a black image, and paints the maze in memory by taking every turn, illuminating where all the walls are.
Areas of high traffic would get higher weights, and areas where no one has ever been get 0 weight. Higher traffic areas are have higher resolution. You would essentially end up with a map of everywhere anyone has been and use it as a prediction model.
I wouldn't be surprised if you could determine what seat a person took in a theater using this method. Given enough users going to the theater, and enough resolution, you would have data mapping each row of the theater, and how wide each row is. The more people visit a location, the higher fidelity with which you could predict that that person is located.
Also, I highly recommend you get a (free) subscription to GPS World magazine if you're interested in the current research into this sort of stuff. Every month I geek out with it.
I'm not sure how great your offset is, because you forgot to include units. ("Around 10 on each axis" doesn't say much. :P) That said, it's still likely due to inaccuracy in the hardware.
The accelerometer is fine for things like determining the phone's orientation relative to gravity, or detecting gestures (shaking or bumping the phone, etc.)
However, trying to do dead reckoning using the accelerometer is going to subject you to a lot of compound error. The accelerometer would need to be insanely accurate otherwise, and this isn't a common use case, so I doubt hardware manufacturers are optimizing for it.
Android accelerometer is digital, it samples acceleration using the same number of "buckets", lets say there are 256 buckets and the accelerometer is capable of sensing from -2g to +2g. This means that your output would be quantized in terms of these "buckets" and would be jumping around some set of values.
To calibrate an android accelerometer, you need to sample a lot more than 1000 points and find the "mode" around which the accelerometer is fluctuating. Then find the number of digital points by how much the output fluctuates and use that for your filtering.
I recommend Kalman filtering once you get the mode and +/- fluctuation.
I realise this is quite old, but the issue at hand is not addressed in ANY of the answers given.
What you are seeing is the linear acceleration of the device including the effect of gravity. If you lay the phone on a flat surface the sensor will report the acceleration due to gravity which is approximately 9.80665 m/s2, hence giving the 10 you are seeing. The sensors are inaccurate, but they are not THAT inaccurate! See here for some useful links and information about the sensor you may be after.
You are making the assumption that the accelerometer readings in the X and Y directions, which in this case is entirely hardware noise, would form a normal distribution around your average. Apparently that is not the case.
One thing you can try is to plot these values on a graph and see whether any pattern emerges. If not then the noise is statistically random and cannot be calibrated against--at least for your particular phone hardware.