Problem Statement: While general activity tracking (e.g., running or walking) is well-established, fine-grained recognition of strength training exercises remains a significant challenge due to complex movements and frequent transitions. Current wearables often require manual input to start sets or struggle to distinguish between exercises that have similar movement patterns, such as bench presses versus shoulder presses. Furthermore, high intra- and inter-individual variability in execution style makes it difficult for standard "one-size-fits-all" systems to provide accurate results without specific adaptations.
Project Goal: The goal is to design and implement an end-to-end Human Activity Recognition (HAR) Pipeline that automatically segments and classifies strength training exercises using data from a wrist-worn wearable. The project aims to investigate if simpler, more interpretable machine learning approaches can effectively handle complex workout environments and whether specialized individual training significantly outperforms generalized models.
Core Tasks:
- Data Acquisition & Preprocessing: Implement a data collection system (e.g., using Bangle.js or similar) to record 3D-accelerometer and barometer data at a sampling rate of at least 10Hz. Develop a pipeline to handle noise, normalize timestamps, and extract dynamic movement components by removing gravitational influence via high-pass filtering.
- Feature Engineering: Design and implement a multi-domain feature vector. This must include time-domain statistics (MAD, Zero-Crossings), frequency-domain metrics (band energy, dominant frequency), and orientation-based features (pitch/roll) to distinguish between spatially similar movements.
- Classification & Post-Processing: Train classification models (e.g., Logistic Regression or Random Forest) to predict exercises per window. Crucially, implement a Post-Processing Logic (e.g., State Machine Decoder, Exercise Gate) to merge window predictions into stable, plausible workout segments and rest periods.
- System Integration: Build a prototype application (Web or Mobile) that allows users to upload raw sensor data and receive a summarized workout report, including recognized exercises, sets, and durations.
Research Focus:
The research focus lies on a formal System Validation comparing model performances. The student is expected to conduct an evaluation using data from different users to calculate standard metrics (Precision, Recall, F1-Score) for both generalized and personalized approaches. Additionally, the student must analyze the Segmentation Quality—evaluating how accurately the system identifies the start and end of sets compared to a ground-truth log.
Technologies: Python (Scikit-learn / Pandas); Embedded JavaScript (Bangle.js) or Mobile Sensors; React / TypeScript; Signal Processing; Machine Learning
Tags: Human Activity Recognition (HAR); Wearable Computing; Machine Learning; Signal Processing; Personalization; Digital Health; Kalchgruber
