What is the real accuracy of telephone accelerometers when used for positioning?

Question

I am working on an application where I would like to track the position of a mobile user inside a building where GPS is not available. The user launches in a well-known fixed location (accurate to 5 centimeters), after which the accelerometer in the phone must be activated to track further movements relative to this fixed location. My question is that in modern generations of smartphones (iphones, android phones, etc.), How accurately can one expect to track the position of some bodies based on the accelerometer on which these phones are usually equipped?

Specific examples would be good, for example, "If I translate 50 meters X from the starting point, 35 meters Y from the starting point and 5 meters Z from the starting point, I can expect my location to be close to +/- 80 centimeters on most modern smartphones "or something else.

I only have a superficial understanding of methods, such as Kalman filters, for fixing drift, although if such methods are relevant to my application and someone wants to describe the quality of the corrections that I could get from such methods, this would be a plus .

Answer 1

If you integrate the accelerometer values ​​twice, you get a position, but the error is terrible. In practice, this is useless.

Here is an explanation of why (Google Tech Talk) at 23:20.

I answered a similar question.

Answer 2

I don’t know if this thread is open or even if you are still trying to use this approach, but I could at least contribute to this, given that I tried the same thing.

As Ali said .... it's awful! The smallest measurement error in accelerometers is ambiguous after double integration. And because of the constant increase and decrease in acceleration when walking (with each step), in fact, this error quickly accumulates over time.

Sorry for the bad news. I also did not want to believe it until I tried to do it ... filtering unwanted measurements also does not work.

I have a different approach, perhaps plausible, if you are interested in continuing your project. (the approach I followed the dissertation at my computer level) ... through image processing!

You basically follow the theory for optical mice. Optical flow, or what is called a view, Ego-Motion. Image processing algorithms implemented in Androids NDK. Even implemented OpenCV via NDK simplifies the algorithms. You convert images to shades of gray (compensate for different illumination values), then set a threshold value, improve the image on the images (to compensate for blurriness of images while walking), then detect the angle (increase accuracy to evaluate the overall result), then match with the pattern, does the actual comparison between image frames and the actual offset of the estimates in the number of pixels.

Then you go through the trial version and the error to evaluate how many pixels represent what distance, and multiply by this value to convert the pixel offset to the actual movement. This works up to a certain speed, although the real problem is that the camera images are still too blurry for accurate comparisons due to walking. This can be improved by setting the camera shutters or ISO (I still play with this).

So hope this helps ... otherwise google for Egomotion for real-time applications. In the end, you get the right things and find out what I just explained. enjoy :)

Answer 3

The optical approach is good, but OpenCV provides several features. Then you get a match (OpenCV provides this).

Without a second reference point (2 cameras), you cannot restore where you are directly because of depth. In the best case scenario, you can estimate depth as a point, take a movement, enumerate an assumption based on several frames, and rethink each depth and movement until it makes sense. Which is not so difficult to code, but it is unstable, small movements of things on the stage screw it. I tried:)

With a second camera, though, it's not that difficult. But cell phones do not.