Abstract
Human Activity Recognition (HAR) using wearable sensors often suffers from a "generalization gap" where models trained on a general population fail to perform accurately for specific individuals. This paper introduces Transfer Convolutional (TrC), a transfer learning framework designed to personalize HAR models. By freezing feature-extraction layers of a pre-trained 1D Convolutional Neural Network (CNN) and fine-tuning only the classification layer with minimal labeled samples, the TrC method achieves superior accuracy compared to traditional machine learning classifiers.
Introduction
The effectiveness of Human Activity Recognition (HAR) systems is frequently hampered by the unique movement patterns and physical characteristics of different users. Traditional approaches to bridge this "generalization gap" often require extensive data collection and manual feature engineering for every new user, which is impractical for consumer applications. This research proposes a system that learns universal movement features from a large population and adapts to specific individuals with minimal human effort.
Methods
The proposed framework utilizes a specialized neural network architecture for sensory time-series data:
- Preprocessing and Embedding: Raw 3D accelerometer signals are discretized to reduce noise sensitivity. An embedding layer then creates a compact representation, mitigating sensor calibration differences.
- Feature Extraction: A stack of 1D Convolutional layers automatically learns local dependencies and scale-invariant features, bypassing the need for manual feature engineering.
- Transfer Convolutional (TrC) Strategy: The model is pre-trained on a large source domain. For a new user, all layers except the final classification layer are frozen. The top layer is then fine-tuned using as few as three labeled instances per activity.
The approach was evaluated using the Sport and Daily Activity (SDA) and WISDM datasets, which provide diverse sensor data and user activities for robust cross-validation.
Key Results
The TrC method was compared against five standard classifiers: Decision Trees (DT), Logistic Regression (LR), Random Forest (RF), Support Vector Machines (SVM), and Quadratic Discriminant Analysis (QDA).
As shown in the figure above, the TrC method consistently outperformed all traditional shallow classifiers. In the SDA dataset, while models like SVM and RF achieved 60-80% accuracy, TrC reached approximately 90%. A similar trend was observed in the WISDM dataset, where TrC maintained a significant lead even when compared to models trained on combined datasets. These results highlight that early CNN layers capture universal movement "motifs," and minimal supervision (three instances per activity) is sufficient for effective personalization.
Conclusion
The Transfer Convolutional (TrC) framework offers a "plug-and-play" solution for personalized HAR. Future research will explore deeper architectures, dynamic freezing strategies, and the computational efficiency of these models on low-power wearable hardware.
Summary updated