Sampling Frequency Estimating of Human Activity Acceleration Data using Transformer-based Regression Model

Your smartphone, smartwatch, and various devices around you are equipped with 3-axis accelerometers. 3-axis acceleration data collected from these devices are used in various fields, and processing acceleration data requires information on the data acquisition status.

However, there is a possibility that the timestamp may be missing or may have been falsified by a malicious person. Therefore, it is necessary to check the sampling frequency when using sensor data. I am working on a method to estimate from 3-axis acceleration data only, at what sampling frequency the data was collected. I have tried various methods and am currently experimenting with a method to predict sampling frequency by building a regression model based on the Encoder portion of the Transformer using histograms generated from acceleration data.

However, while the method manages to produce good prediction accuracy if the data are from the same domain, the errors become large if the data are from different domains. Prediction accuracy is also low between data acquired by different exercises, such as walking and running activities. In addition, it is very difficult to estimate the sampling frequency because the theme itself is abstract and the acceleration data itself takes various values depending on the type of motion. It is difficult to manage my time because I have many other things to do, such as job hunting and part-time jobs, and I also want to cherish the time I have for my hobbies. However, I am advancing my research little by little through friendly competition with my colleagues and professors, and I am having a lot of fun.

This is link of my paper → https://dl.acm.org/doi/10.1145/3594739.3610794

This is my e-mail address → hinase.kawano@iis.ise.ritsumei.ac.jp

Any advice would be appreciated!