TL;DR
This paper presents a novel approach to real-time human activity recognition (HAR) on smartphones using lightweight and accurate features derived from accelerometer data. By analyzing acceleration signals as an autoregressive (AR) process and employing Kernel Discriminant Analysis (KDA) to handle position-variability, the system achieves high accuracy with low sampling rates and short data windows, significantly improving battery life and responsiveness for mobile HAR.
What Problem Does This Paper Address?
Human Activity Recognition (HAR) using smartphones is a growing field with applications ranging from healthcare to personalized services. However, existing online HAR systems often compromise battery life and system responsiveness by requiring high sampling rates or long data windows to achieve acceptable accuracy. Furthermore, many systems are sensitive to the phone's position on the body, leading to high within-class variance and reduced accuracy. This paper tackles these challenges by developing a real-time, subject-independent HAR system that uses lightweight features, operates efficiently on a smartphone, and is robust to phone placement variations.
Methods
The core of this research lies in its feature extraction and dimensionality reduction techniques:
- Exploratory Data Analysis (EDA): The authors performed EDA on accelerometer data to understand the underlying signal characteristics. They used autocorrelation and lag plots to demonstrate that activity acceleration signals can be modeled as an autoregressive (AR) process.
- Autoregressive (AR) Modeling: Based on the EDA findings, AR coefficients were chosen as features. AR models predict current signal values from past values, offering a compact representation of the signal's dynamics. The optimal AR model order (10) and window length (3 seconds) were determined using Akaike Information Criteria (AIC) and Signal-to-Noise Ratio (SNR), respectively.
- Noise Reduction: A moving average filter of order 3 was applied to mitigate outliers identified during EDA.
- Kernel Discriminant Analysis (KDA): To address the issue of high within-class variance caused by different phone positions, KDA was employed. KDA is a non-linear dimensionality reduction technique that aims to find discriminating features in a transformed feature space, making the system more robust to variations in phone placement.
- Artificial Neural Networks (ANNs): Standard feed-forward, backpropagation ANNs were used as classifiers, chosen for their superior performance in previous studies.
Datasets
The study utilized data collected from 30 healthy subjects (18 males, 12 females, aged 26-35) performing six common activities: standing, walking, walking-upstairs, walking-downstairs, running, and hopping. Data was collected using LG Nexus 4 smartphones at various sampling rates (20 Hz to 120 Hz) and from five different phone positions (left/right front pockets, left/right back pockets, and jacket's inner pocket). For evaluation, 10 new subjects (6 males, 4 females) from different age groups and with varying physical characteristics were recruited for subject-independent testing.
Key Insights and Ideas
- AR Process for HAR: The finding that activity acceleration signals can be effectively modeled as an autoregressive (AR) process is crucial. This allows for the extraction of lightweight and informative features (AR coefficients) that are computationally efficient.
- Low Sampling Rate and Short Window Viability: The research demonstrates that high accuracy can be achieved with a low sampling rate (20 Hz) and a short data window (3 seconds), directly addressing the battery life and responsiveness concerns of mobile HAR.
- KDA for Position Invariance: The successful application of KDA to mitigate the impact of phone placement variations is a significant contribution. This makes the system more practical for real-world scenarios where users are unlikely to maintain a fixed phone position.
- Subject-Independent Evaluation: The rigorous subject-independent evaluation with new participants from diverse demographics strengthens the generalizability of the findings.
Key Results
The proposed AR-feature-based system, combined with KDA and ANNs, achieved a mean recognition accuracy of 87.1% in a subject-independent setting using a low sampling rate (20 Hz) and a short window (3 s). This performance significantly outperformed other commonly used features:
- AR-features: 87.1%
- Time Domain (TD) features: 67.58%
- Discrete Cosine Transform (DCT) features: 48.55%
- Fast Fourier Transform (FFT) features: 45.41%
The AR features showed a minimal drop in accuracy (around 3%) with increasing sampling rates, whereas FFT and DCT features required significantly higher sampling rates (around 100 Hz) to achieve comparable, though still lower, accuracy. Energy consumption analysis also indicated that AR features are more energy-efficient for HAR on smartphones.
Limitations and Future Work
- Real-life Scenario Idealization: While the system handles multiple positions, it's acknowledged that a truly ideal real-life system would require position/orientation-independent features. The authors hypothesize that projecting data onto a standard/global reference coordinate space could be a future direction.
- Increasing Accuracy: Despite achieving reasonable accuracy, further improvements are desired. The authors suggest exploring the use of data collected during the "usage-state" (when a new user is actively using the phone with a trained classifier) to adapt and improve recognition rates for that specific user.
- Computational Complexity: Implementing advanced approaches for position independence might increase computational complexity, which is a challenge for resource-constrained mobile devices. Future work will investigate this.
Summary updated