TL;DR
This paper introduces a novel user-adaptive algorithm for human activity recognition that personalizes models for individual users without requiring manual data labeling. By combining K-Means clustering for automatic annotation, Local Outlier Factor (LOF) for robust sample selection, and Multivariate Gaussian Distribution (MGD) for classification, the system achieves high accuracy in recognizing activities like walking, running, and falls, while significantly reducing computational complexity during testing.
What Problem Does This Paper Address?
Human activity recognition (HAR) systems, particularly those using wearable sensors, often struggle with personalization. Data distributions from inertial sensors vary significantly between users, meaning a model trained on one person's data performs poorly when applied to another. This "cross-person error" limits the practical application of HAR in areas like healthcare and personalized recommendations. While some approaches attempt personalization, they often rely on manual data labeling, which is impractical and time-consuming. This paper addresses the need for an automated, user-adaptive HAR system that can build accurate, personalized models without user intervention.
Methods
The proposed user-adaptive algorithm consists of several key stages:
- Data Preprocessing and Feature Extraction: Raw accelerometer and gyroscope data are filtered and segmented into windows. Features are extracted from these windows, focusing on time-domain features like mean, standard deviation, energy, and synthesized acceleration/angular velocity magnitudes. A "tilt angle" feature is also computed.
- Automatic Annotation (K-Means Clustering): A novel K-Means initialization method is employed. Instead of random initialization, it uses centroids from a pre-collected, labeled dataset that is similar in distribution to the target user's data. This allows for automatic clustering and labeling of the new user's data.
- High Confidence Samples Selection (Label-based LOF): To ensure robust model training, a label-based Local Outlier Factor (LLOF) is used. This technique identifies and removes low-confidence samples (outliers) within each cluster based on their relative density, ensuring that only reliable data points are used for personalization.
- Model Classification (Multivariate Gaussian Distribution - MGD): Three MGD models are trained, one for each of the non-fall activity categories: Light Intensity Activity (LIA), Moderate Intensity Activity (MIA), and Vigorous Intensity Activity (VIA). These models estimate the probability of a new data point belonging to each activity class.
- Fall Detection: Falls are initially grouped with VIA due to their high signal variance. Specific thresholds on the MGD output for VIA and the "tilt angle" feature are then used to distinguish between VIA and actual falls.
Datasets
The paper describes experiments conducted with data from 10 participants.
- Training Data: A pre-collected, labeled dataset was used to initialize K-Means centroids. For training the personalized MGD models, a portion (4:1 ratio) of each individual participant's collected data was used.
- Testing Data: The remaining portion of each participant's collected data was used for testing. This included activities from all categories (LIA, MIA, VIA, and simulated falls).
- Benchmark Comparison: The proposed method is compared against existing algorithms from literature, specifically those by Viet et al. [29], Deng et al. [31] (for personalization), and Tong et al. [22], Shi et al. [27] (for fall detection).
Key Insights and Ideas
- Automated Personalization is Key: The core insight is that effective HAR requires personalization, but manual labeling is a significant bottleneck. The proposed K-Means initialization method offers a practical solution for automatic annotation.
- Relative Density Matters for Sample Selection: Simply selecting "high confidence" samples isn't enough. Using LOF to consider relative density helps to filter out noisy or ambiguous data points, leading to more robust personalization.
- MGD for Efficient Classification: MGD provides a computationally efficient way to model activity probabilities, making it suitable for resource-constrained wearable devices, especially when dealing with low-dimensional feature spaces.
- Fall Detection as Anomaly Detection: The approach cleverly handles the scarcity of real fall data by treating fall detection as an anomaly detection problem, leveraging the MGD models trained on normal activities.
Key Results
- High Accuracy: The personalized algorithm achieved high F-measure scores across all four categories: LIA (0.9875), MIA (0.9729), VIA (0.9692), and Fall (0.9766) on average.
- Significant Improvement over Generic Models: Personalization led to a statistically significant improvement in F-measure compared to generic, non-personalized models, particularly for MIA, VIA, and fall detection.
- Computational Efficiency: The proposed method demonstrates significantly lower computational complexity in the testing phase (0.02s) compared to existing methods (e.g., 2.83s, 0.15s), making it highly suitable for real-time applications.
- Robustness to BMI: The study noted that BMI did not significantly impact the effectiveness of personalization, with both normal and abnormal BMI groups showing significant improvements.
Limitations and Future Work
- Limitations:
- The effectiveness of the K-Means initialization relies on the availability of a pre-collected, labeled dataset with a similar distribution to the target user's data.
- While fall detection performance is competitive, it's not the absolute best compared to some specialized fall detection algorithms.
- Future Work:
- The authors plan to explore multi-sensor fusion techniques to enhance activity recognition.
- They aim to apply deep learning methods for detecting more complex abnormal behaviors.
- Further testing with genuine fall datasets is planned.
Summary updated