Detecting the movement direction using an accelerometer

[marcospassos]

I'm a few days trying to get the direction of movement using an accelerometer. For my application I need to put an object on the table (z axis fixed) and detect the movements. I know perhaps the accelerometer is not the best sensor for this kind of application but is the sensor that is available on the devices that I need to use.

For example:

- I move it to the right, left, right and then left quickly.
- The expected information: right, left, right, left

The pattern for the movement start is simple: I get the signal and compute the difference between the current value and last value (ignoring small values) and if is negative, than it is moving to the left, if is positive it is moving to the right. So, I've created a stop condition. When I've obtained X numbers of zeros of the difference I understand it is stoped. This algorithm works great and is very accurate.

The big problem is when I'm movig to right and I quickly move to the left. The algorithm cannot detect the change because he is waiting the stop and it doesn't happens.

I've analyzed the data several times, but I can not find a pattern to the end of the movement in order to create an new algorithm.

Bellow are some graphics that I've generated using the data obtained from accelerometer.

http://img31.imageshack.us/img31/9544/graphsjk.jpg

The chart on the left side is a simple and slow movement from the right to the left. In the right side is the problematic situation where the movement is from the right to the left and to the right again.

I would be grateful if you could help me.

Thanks in advance,
Marcos

 

[A.T.]

Without a gyroscope, this will work only if the orientation (all axes) of the device is fixed. You have to know the initial velocity of the object when you start recording the acceleration (presumably zero, if the user is advised to hold the device still for a moment). Then you have to integrate the acceleration to get the velocity. There should be plenty of sample code for this online in any language used on mobile devices.

The pattern for the movement start is simple: I get the signal and compute the difference between the current value and last value (ignoring small values) and if is negative, than it is moving to the left, if is positive it is moving to the right.

That sounds like differentiation. What you have to do is something like this:

current_velocity = velocity_dt_ago + (acceleration_dt_ago + current_acceleration) / (2 * dt)

There a better integration algorithms though. Google for "runge kutta 4".

 

[marcospassos]

Hello A.T, thanks so much for your reply.

The velocity is not important to me. I just want to know the direction. In fact, I have the direction using my algorithm but it doesn't work when I change quickly. I believe there is a formula or pattern to identify when it occurs through the graph peaks.

http://img403.imageshack.us/img403/831/graphs2.jpg

Observing the following graph I can suggest that at point 1 the object starts moving to the left, because the graphic was stable and started decreasing. At point 2, I believe the movement from the left definately has stopped and started moving to the right, once the values started increasing. At point 3 the object is stopped again.

As it is possible recognize looking at, I believe is possible create a algorithm for this, using some formula or pattern.

Am I wrong?

 

[A.T.]

The velocity is not important to me. I just want to know the direction.

Velocity is a vector. It tells you the direction as well. You have to apply the intergation to all three components (x,y,z).

I believe is possible create a algorithm for this, using some formula or pattern.

Yes, integration.

 

[marcospassos]

A.T,

I'm trying get it work. What I'm doing:

- First, the algorithm "calibrates" the sensor getting 50 samples and using the average as the offset
- Subtracts the offset from the accelerometer value
- Stores the current value of accelerometer
- Makes comparisons using the speed in order to detect the direction

I think I'm doing something wrong, because doesn't works well. I moved the device from the left to right and then I stoped, it shows: left, right, stoped. Sometimes it shows always right or left.

Can you help me?

My algorithm:

// calibration... offsetX = SUM(50 samples)/50

accX[1] = (x - offsetX);

// Mechanical Filtering (remove some noise)
if((accX[1] < 0.2) && (accX[1] > -0.2)) {
accX[1] = 0;
}

// First X integration:
velocityX[1] = velocityX[0] + accX[0] + (accX[1] -accX[0]);
// Second X integration:
positionX[1] = positionX[0] + velocityX[0] + (velocityX[1] - velocityX[0]);

accX[0] = accX[1];
velocityX[0] = velocityX[1];

if(velocityX[0] < 0)
debug = "left";
else if(velocityX[0] > 0) {
debug = "right" ;
} else {
debug = "stoped";
}
​

 

[A.T.]

// First X integration:
velocityX[1] = velocityX[0] + accX[0] + (accX[1] -accX[0]);
// Second X integration:
positionX[1] = positionX[0] + velocityX[0] + (velocityX[1] - velocityX[0]);

Ahhm... you are adding and subtracting the same values there (accX[0] and velocityX[0]).

Simple intergation works like this:

velocityX[0] = 0 // we assume the device was static when we started recording acc
positionX[0] = 0

for every sample i > 0:
    dt = time[i-1] - time[i] // if sampling rate is constant simply use 1 / frequency here
    velocityX[i] = velocityX[i-1] + (accX[i] + accX[i-1]) / (2 * dt)   
    positionX[i] = positionX[i-1] + (velocityX[i] + velocityX[i-1]) / (2 * dt)