Bachelor Thesis: Benchmarking Time-Frequency Transforms for Deep Learning-Based Human Activity Recognition

Background

Human Activity Recognition (HAR) using data from wearable or smartphone sensors (e.g., accelerometers, gyroscopes) is an increasingly active research area with broad applications in healthcare, sports, mobile computing, and human-computer interaction. While many deep learning approaches operate directly on raw time-series data, this can constrain their ability to capture frequency-dependent patterns that are often crucial for distinguishing among different activities.

To address this limitation, sensor signals can be transformed into the time-frequency domain using techniques such as the Short-Time Fourier Transform (STFT) or Wavelet Transforms (CWT/DWT). These methods yield 2D representations, spectrograms or scalograms, that simultaneously preserve temporal and spectral characteristics of the signal. Such enriched representations are particularly well-suited for Convolutional Neural Networks (CNNs), which excel at identifying spatial and hierarchical patterns in 2D data. Leveraging CNNs in this context enables more effective learning of complex activity signatures.

Although various time-frequency transforms have been used in HAR research, a thorough evaluation comparing their theoretical and practical advantages for HAR, as well as their compatibility with CNNs, classification performance, and computational efficiency, is lacking. This thesis aims to fill that gap by systematically analyzing and benchmarking a range of time-frequency transforms across multiple datasets to assess their relative strengths and limitations in the context of human activity recognition.

Tasks

• Review literature on time-frequency transforms (e.g., STFT, CWT) applied to HAR.
• Analyze and compare transforms regarding theoretical and practical suitability for HAR, efficiency, and CNN compatibility.
• Implement and train CNNs on various HAR datasets for activity classification using different time-frequency representations.
• Conduct thorough evaluation of classification performance and computational efficiency.

Requirements

• Proficient programming skills in Python; experience with PyTorch is a plus.
• Solid understanding of deep learning concepts, particularly CNNs.
• Basic knowledge of signal processing is a plus but can be acquired during the thesis.

If you are interested in this topic, please contact Max Burzer (burzer@teco.edu)