TL;DR
This paper presents a privacy-preserving system for monitoring Activities of Daily Living (ADLs) using audio analysis on a low-cost Raspberry Pi. The system focuses on a bathroom scenario, demonstrating that it can accurately recognize specific sounds like flushing or showering without storing or transmitting raw audio, making it suitable for elder care and health monitoring.
What Problem Does This Paper Address?
The increasing elderly population necessitates robust home monitoring solutions to enable independent living. While AI and machine learning offer promising avenues for this, deploying them in commodity electronics faces challenges: 1) privacy concerns with sensitive data collection, and 2) the need for computationally efficient and easily configurable systems that don't require specialized machine learning expertise from technicians. Existing solutions often rely on centralized services, exacerbating privacy risks, or require complex setup. This paper tackles the problem of creating an on-device, privacy-aware ADL monitoring system that is both effective and user-friendly for deployment in real-world home environments.
Methods
The core of the system is a Raspberry Pi single-board computer equipped with a microphone. This device handles both audio acquisition and real-time processing, including feature extraction and classification. The key innovation lies in performing all analysis locally on the Raspberry Pi, ensuring that only abstract Activity of Daily Living (ADL) logs are generated, not raw audio data. This significantly enhances privacy.
Communication between the Raspberry Pi and a remote interface (like an Android app) is managed via the MQTT protocol, a lightweight publish/subscribe messaging protocol ideal for IoT applications. This allows for remote commands (e.g., to start/stop recording) and retrieval of processed results.
A crucial aspect is the system's calibration and training procedure. It's designed to be user-friendly for technicians without machine learning backgrounds. Through a mobile application, technicians can guide the recording of specific sounds (e.g., flushing, showering) and label them. The Raspberry Pi then uses this data to train a Support Vector Machine (SVM) classifier. The system extracts 34 short-term audio features, which are then aggregated into a 68-dimensional feature vector per 1-second audio segment. The paper specifically mentions using probabilistic SVMs for their generalization capabilities and a cross-validation procedure to select optimal parameters.
Datasets
The paper introduces a custom dataset for the bathroom activity monitoring use case. This dataset consists of:
- A large training dataset of over 400 audio segments for "static" training (pre-trained classifiers).
- A subset of 4 continuous audio recordings with ground-truth annotations, totaling 7 minutes, used for both training and testing.
The recordings were made in two different bathroom locations, allowing for evaluation of the system's adaptability to different acoustic environments. The dataset and ground truth are made publicly available.
Key Insights and Ideas
- Privacy-Preserving ADL Monitoring: The most significant insight is the successful demonstration of an on-device audio analysis system that provides ADL logs without compromising user privacy by avoiding raw audio storage or transmission. This is a critical step towards widespread adoption of home monitoring technologies.
- Low-Cost, Edge Computing Solution: Utilizing a Raspberry Pi for both data acquisition and analysis makes the system highly cost-effective and eliminates the need for a dedicated workstation, a significant advantage for home deployment.
- User-Friendly Calibration: The emphasis on a simple, technician-guided calibration process, abstracting away the complexities of machine learning, is a key enabler for practical deployment in diverse home settings.
- Location-Specific Training Matters: The experimental results strongly suggest that training classifiers on audio data specific to the target environment (e.g., a particular bathroom) yields significantly better performance than using generic, pre-trained models, even if those models were trained on much larger datasets.
Key Results
- The system achieves a satisfactory classification performance, with an overall F1-measure of 68.1% in the initial experimental setup (single-recording training).
- When training with three recordings (instead of one), the overall F1-measure improves to 73.8%, demonstrating the benefit of more localized training data.
- The system achieves 2x real-time performance on the Raspberry Pi, confirming its computational efficiency for on-device processing.
- The paper highlights that using "static" models trained on external data leads to poor classification performance (F1 scores around 57-59%), underscoring the value of the proposed localized training approach.
Limitations and Future Work
- Limited Dataset Size: While the custom dataset is valuable, the authors acknowledge that larger datasets would likely improve performance further.
- Single Microphone: The current setup uses a single microphone. Future work could explore using multiple microphones for more robust spatial information.
- Future Directions: The authors propose extending the calibration procedure to incorporate a "base dataset" for initial tuning, rather than retraining from scratch. They also aim to use long-term temporal knowledge to smooth classifier results based on prior event knowledge.
Summary updated