TL;DR
This paper introduces a novel, cost-effective pressure mat system for human activity recognition (HAR) that leverages Hyperdimensional Computing (HDC) for enhanced noise resilience. The system achieves high accuracy in detecting both static and dynamic human activities, significantly outperforming existing methods, especially in the presence of various noise sources, while consuming substantially less energy.
What Problem Does This Paper Address?
Traditional systems for indoor human activity recognition, particularly those using pressure mats, often face challenges with accuracy and robustness. High-resolution pressure mats can be expensive, and the machine learning models used for activity recognition, such as Convolutional Neural Networks (CNNs), are computationally intensive and prone to errors caused by noise. This noise can originate from various sources, including sensor inaccuracies, input data fluctuations, communication interference, and memory errors. The existing solutions struggle to maintain performance in real-world scenarios where these noise factors are prevalent, making them less suitable for resource-constrained environments like smart homes. The paper aims to bridge this gap by developing a system that is both accurate and highly resilient to noise, while remaining cost-effective and energy-efficient.
Methods
The core of this work lies in the integration of a low-cost pressure mat with a novel Hyperdimensional Computing (HDC) classifier.
- Pressure Mat Hardware: The system utilizes a cost-effective pressure mat constructed from Velostat material, a force-sensitive resistor. This mat is equipped with 512 sensing areas, achieved through a grid of carbon tapes. Signal acquisition is managed by multiplexers (MUXes) and analog-to-digital converters (ADCs), with a Raspberry Pi (RPi) handling data processing and learning.
- Novel HDC Encoding: The paper proposes a new window-based encoding method for HDC. Instead of treating an entire image as a single input, this approach divides the input into smaller windows. Each window is then encoded using a random projection technique, preserving local spatial and temporal features. These encoded window hypervectors are then superimposed to create a single hypervector representing the entire input. This window-based encoding is designed to be more robust to noise and to better capture local patterns compared to traditional HDC encoding.
- Noise Resilience: The HDC framework itself is inherently noise-resilient due to its high-dimensional representation. The proposed novel encoding further enhances this resilience by effectively preserving local information, making the system less susceptible to various noise types, including input noise (shifts, blurriness, rotations), sensor noise (Gaussian, white noise), and memory/communication noise (packet loss, bitflips, Gaussian noise).
Datasets
The authors collected two datasets for training and evaluation:
- Static Activities: This dataset captures stationary postures, including "no_press" (no movement on the mat) and "standing still."
- Continuous Activities: This dataset includes dynamic movements such as "jump," "walk," "left/right (lr)_shift," and "tiptoe."
The data consists of 7-frame time series, with each frame represented by 32x16 digital values. The datasets were split into training (80%) and testing (20%) sets.
Key Insights and Ideas
- HDC for Pressure Mats is Novel: This paper marks the first known application of HDC for analyzing pressure mat data, opening up a new avenue for HAR research.
- Window-Based Encoding is Key to Noise Resilience: The proposed window-based encoding significantly improves the system's ability to handle noisy data by focusing on local features, a concept inspired by CNNs but implemented within the HDC framework.
- Cost-Effectiveness and Energy Efficiency: The combination of Velostat for the pressure mat and HDC for classification leads to a system that is not only accurate but also significantly more affordable and energy-efficient than traditional CNN-based approaches.
- Unified Noise Resilience: The system demonstrates robust performance across a wide spectrum of noise types, from sensor inaccuracies to communication errors, highlighting its practical applicability in real-world smart home environments.
Key Results
- Accuracy:
- For static activities, the proposed HDC system achieves accuracy comparable to state-of-the-art CNNs (98.96% vs. 98.88%).
- For continuous activities, the system achieves a significantly higher accuracy of 93.19%, outperforming CNNs (83.72%) by 9.47%.
- In the presence of Gaussian noise, the system maintains an accuracy of 92.15% (97.51% for static data), showing an improvement of 13.19% (8.77%) over state-of-the-art CNNs.
- The system demonstrates remarkable resilience to memory and communication noise, achieving accuracy improvements of up to 33.77% and 18.72% for dynamic and static data, respectively, compared to the best state-of-the-art methods.
- Efficiency:
- The system consumes 85% less energy than CNNs for comparable testing accuracy.
- The proposed method is up to 7.3x faster than CNNs for static data and 1.44x faster for dynamic data in terms of total execution time.
- Dimensionality: The proposed HDC uses compact 200-dimensional vectors, significantly smaller than the typical 2k-10k dimensions used in baseline HDC, contributing to its efficiency.
Limitations and Future Work
- Limitations:
- The paper acknowledges that for more complex tasks and data, higher dimensions might be required for HDC, which could pose challenges for memory and computational bottlenecks.
- The linear nature of HDC encoding has limitations in capturing highly complex image patterns, suggesting that combining HDC with NN-based feature extraction might be beneficial for certain tasks.
- The current study primarily focuses on static data when analyzing noise, with dynamic data noise analysis being a subject for future work due to its complexity.
- Future Work:
- Investigate the use of signals from human pose and shape analysis applied to pressure images to achieve more nuanced HAR.
- Extend the classification capabilities to more complex cognition tasks using heterogeneous data.
- Conduct a comprehensive examination of datasets with imbalanced labels and explore data augmentation techniques for HDC.
- Further study system maintenance in the presence of noise and explore hyperspace sparsification techniques for reduced memory usage.
Summary updated