Driving, Cycling, and Walking Activity Recognition with Accelerometers: How to distinguish between the activities using data obtained from an accelerometer?

I am attempting to design a matlab program that will differentiate between different activities.

The data will be gathered from an accelerometer.

Is there a sensible way to do this?

Many thanks,

[Jan]

Where do you attach the acceleratormeter? The methods differ if you have both feet or just the center of mass.

You find many corresponding articles in the literature for orthopedics. This is not a Matlab problem.

[Star Strider]

This may require a machine learning approach. See for example Automated Feature Engineering for Classification and perhaps other approaches as well. (This is not an area of my specific expertise, so start here and explore.)

It may also be appropriate for a signal processing approach, such that using the stft could be used to characterise the frequency spectra of various activities, and then classify using them. I have some relevant experience with that, since I once used it to determine (with reasonably accuracy) a laboratory cognitive task a particular subject was doing using STFT spectra to detect patterns of electroencephalogram (EEG) activation. (We used a simple linear discriminant for that decades ago. The nonlinear classification approaches now available in MATLAB could have made it much more robust. It was published in a respected refereed journal.) First, collect some representative signals and then see what they have in common and how they differ. The principal component analysis (pca) can help with that. Then,go from there!

[ahmad Mohammadi]

@Jan Thanks for your reply. It is actaully from phone accelerometer. Matlab has an ability to collect such a data and do analisys.

[ahmad Mohammadi]

@Star Strider Thanks for the information. Actually, I am looking for the question whether somebody applied such a coding or not and start from there.

[Star Strider]

My pleasure!

That is the approach I took in my research. I knew what the tasks were, and split the data into training and testing and compared the classification accuracy. Start with data you know, and then determine the relevant features of each and how they differ from the others.

You might be able to get some of the activity type information from the GPS data (location, veolcity, etc.) as well as the accelerometer information. That could help label it.

Answer 1

[Cris LaPierre]

I would point you to the following examples

• Human Activity Recognition Simulink Model for Smartphone Deployment
• Human Activity Recognition Using Signal Feature Extraction and Machine Learning
• Human Action Recognition using KTH Dataset
• Summary Statistics as Features (comes from our Data Processing and Feature Engineering with MATLAB course on Coursera)