I need some help infact a lot of help undersating the autocalibration procedure of physcial activity acceleromters. I have been reading a paper by van hees et al, JAP 2014. but struggling to reproduce the steps as I am not a very technical/mathematical person. Does anyone know simplified steps of removing the effect of gravity from accelerometer data.
Thank you very much.
Saima
Dear Saima
It seems to me that milligals too much precision for You. The article discusses the calibration of high-precision systems. A gravity in any case You can't remove using only accelerometers on the mobile system. It need additional information: vertical, gyroscope or something else.
Best regards.
Thats in reference to a global position/reference.
Depending on where the device is attached you could subtract the offset due to gravity while they are standing still, i.e. for the vertical acceleration +1 to the -1g orientation. This is rough/quick. (been used many times for a device attached to the trunk).
The paper by Moe-Nilssen uses trigonometric functions to correct each axis but relates to correcting during gait/walking with a device on the lower back. "A new method for evaluating motor control in gait under real-life environment conditions. Part 1: the instrument" Clinical Biomechanics 1998, 13, 320-327.
Bench testing: orientating the device so "each individual" axis is sensitive to gravity and subtracting/adding the mean value from each orientation to each respective axis is another crude way.
I wrote the paper you are referring to. Autocalibration has nothing to do with removal of the gravitational component from an acceleration signal. You are confusing two very different topics. If your question is about the removal of gravitation component then you may want to check out my 2013 PLoSONE paper: http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0061691 in addition to the suggestions by Alan
Thanks a lot dear Vincent for answering.
So I didn't get the paper properly. How was the calibration done then!
Also removing the gravitational component has an impact on the accelerometer data, right? Is it advised to remove it or not?
Thanks
Saima
The autocalibration is done by looking at the average acceleration per sensor axis during periods of no movement and comparing these values with what value you would theoretically expect given the fact that the three axes should always combine to 1 gravity when there is no movement.
The method we demonstrate in the paper uses real life data for doing this, by which there is no need for additional experiments. The R code of how it is done is included in the paper and part of R package GGIR. The paper also provides an explanation of all the steps in the algorithm. In recent years other groups have copied this algorithm and implemented it in Python, Matlab and C (N Hammerla and colleagues in Newcastle) and Java (A Doherty and colleagues in Oxford).
For gait research / lab research the problem may be that there are not enough data sections in your acceleration data with no movement, so either you would have to explicitly plan no movement periods in your study protocol with a variety of sensor orientations or rely on real life data collected with the same accelerometer serial number in a separate experiment. The latter has the disadvantage that the calibration error may differ between experiments and by that also the correction coefficients.
Saima, my answer to your second question is:
The gravitational component, when separated, can be informative about the orientation of the accelerometer and by that about the orientation of the body segment to which it is attached. In our 2013 paper we demonstrated that not attempting to remove the gravitational component reduces the ability to explain variation in activity-related energy expenditure, see results for metric EN. Therefore, separating the gravitational and movement components seems a sensible thing to do. Depending on your application you would then use one of them or both. Not everyone uses the same computational approach to this though, but in general I think people agree that attempting to separate the components is best practise... even traditional count accelerometers do this by applying a frequency filter.
Dear Vincent and Soren,
I am extremely grateful to you for making things much clear. Just to tell you a little about my data: My participants wore Actigrah GT3X+ accelerometer for 7 days (including night time) during their daily activities. So now I am cleaning the data to get the total volume of physical activity per week and see the relationship with metabolic factors (blood samples analysis). Main challenge for me is the data processing from the accelerometer (wear time, non wear time and actual physical actvity). It does have its own software Actilife to get outcomes like time in differnt categories and kilo calories per week etc. But I am confused as ti which is the best method to use. I wanted to use raw acceleration instead of the software processed data.
I will give a detailed read to the article you suggested.
Thank you very much
Saima
Note that R package GGIR which I developed can handle Actigraph GT3X+ data if you export it to csv format. So, maybe this offers an alternative processing pipeline or at the very least an additional reference point for designing your own analyses.
- Badges
- Science topic
- Similar topics
- Control Systems Engineering
- Filtering