University of Twente Student Theses


Characterizing types of convolution in CNN for step detection using IMU data

Yin, L. (2022) Characterizing types of convolution in CNN for step detection using IMU data.

[img] PDF
Abstract:Step detection is essential to realizing a well-functioned pedometer-enabled application. Recent works of IMU-based step detection have shown that Convolutional Neural Network (CNN or ConvNet) is promising in dealing with sequential time series data and results in good step detection accuracy. However, no related works have been done to compare different convolution types used in the CNN model for step detection. This paper investigated and compared the performance (e.g., model accuracy and complexity) of various convolution types (e.g.,1D, 2D, 3D) in a standard shallow CNN model for step detection. We conducted the experiments using only accelerometer data under regular walking mode from a public IMU dataset that contains records of 30 participants under different walking scenarios. Results showed that the 1D convolution type has relatively higher accuracy than 2D and 3D convolutions, and under a 2D convolution context, the depthwise separable convolution type generated the lowest model complexity.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Business & IT BSc (56066)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page