hi, I would like to calculate the distance of a step with an accelerometer 6 axis 3 accelero 3 gyro attach to the belt of the human . When I integrate, I'm losing too much accuracy.
I see some research about a way to determine the distance of steeps with neural network models. Does anyone have an idea if it is really possible and is this method precise ?
If yes, do someone know the type of the model and the input parameters ?
If someone has another idea of who could I have the distance of one step with an accelerometer ?
Thanks
Hello Clara Sergeraett,
Evidently, you cannot integrate the signal directly, because you will integrate also the error, you should try a Kalman filter, maybe will help you obtain one model.
In the other hand, I recommend you to use a RNN (recurrent neural network) in python. The RNN works with time series data, and you can set as input a signal with x lenght. For doing so, you should gather some test examples for training and test set.
Best regards,
Juan Zuluaga
Hello Juan Zuluaga
thanks for your answer it is really interesting.
- Could you explain me why do you choose the RNN as network?
- Also, you say that as input I have one signal. With my accelerometer I have 6 signals differently. Do I need to give the 6 signals without any treatment or do I need to change them?
- Last question, you talk about the Kalman filter effectively I think about that, but if I understand correctly how Kalman filter function, I need to compare 2 signals which give the same thing for example accelerometer and GPS signal for speed. At which other signal do you think?
Best regards,
Clara
Hello Clara,
The RNN, have the following advantage in comparison with other types of ANN, after training the RNN, you could change the dimension of the input vector (who will have the features or the values, in this case 6 different signals). I though I've chosen the RNN because fits the time series data, anyway you may try other types of models. for example, you could look in the Kaggle webpage and look for a "kernel" that suits you, if you want more information, see this:
https://www.kaggle.com/tigurius/recuplots-and-cnns-for-time-series-classification
On the other hand, if you want to expand your knowledge or explore a little bit more about RNN and other type of architectures, you should take a look to this book:
"Hands-On Machine Learning with Scikit-Learn & TensorFlow" It helped my a lot.
Best regards,
Juan
Dear Clara Sergeraett,
the answer at your question you will find in the attached paper.
Regards, A. Tunik.
Dear Anatol Azarievich Tunik,
Your research is really interesting. thanks a lot to share it with me.
In my problem, I have only the acceleration in one point following the axes x,y,z. The goal of my research is to determined the humans' steps length without the GPS to have the velocity and without integrate the acceleration. So the goal is to determined the velocity or the steps length with only the accelerometer and the angular velocity.
Could you confirm me that in your research, you use GPS to have the velocity. if yes, do you think it's possible to do the same thing without the GPS ?
if I understand in your research, you improve the signal of the acceleration, with that you can integrate it and have the angular velocity ?
thanks for your answer
Regards, Clara
Usage of any inertial sensors requires external correction (GPS in our case), otherwise the integration errors will be unconstrained due to the sensor's biases, which are immanent to this kind of sensors. Generally speaking, we really
"improve the signal of the acceleration" by GPS. From my point of view, solution of your problem doesn't exist without external correction.
Regards, Anatol Tunik.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs