TL;DR
This paper proposes a cost-effective occupancy detection system using only audio signals, achieving high accuracy for presence detection (up to 99%) and respectable accuracy for head counting (up to 95%) in commercial and residential environments. By carefully selecting relevant audio features and employing a simple decision tree classifier, the system offers a privacy-preserving and computationally efficient alternative to more complex multi-modal solutions.
What Problem Does This Paper Address?
Accurate occupancy detection, encompassing both knowing if someone is present and how many people are in a space, is crucial for smart buildings. It enables energy savings (e.g., optimizing HVAC), enhances safety (e.g., emergency evacuations), and improves user comfort. However, real-world deployment faces challenges: varying environments, background noise, the need for privacy (making cameras less ideal), and the trade-off between sensor cost/complexity and accuracy. Existing solutions often rely on single modalities with limitations or complex sensor fusion, which can be expensive and computationally demanding. This paper aims to address these challenges by developing an affordable and effective occupancy detection system using only audio signals, a modality that is readily available, privacy-friendly, and offers broad coverage.
Methods
The authors propose a multi-stage methodology for audio-based occupancy detection:
- Feature Extraction: They extract a comprehensive set of low-level descriptive audio features (LLDs) using a short frame length (25 ms) with no overlap. These include features like Zero Crossing Rate, Energy, Brightness, Spectral Variance, MFCCs (and their derivatives), and others.
- Feature Segmentation & Functionals: Audio features are then segmented into fixed time windows. For each segment, various "functionals" (e.g., Mean, Median, Standard Deviation, Absolute Integral, Dynamic Range) are applied to capture the temporal evolution of the LLDs. This step aims to extract more meaningful information from longer time spans.
- Feature Selection: Recognizing the vast number of potential audio features, the authors employ two feature selection methods:
- Univariate Chi2: This method ranks features based on their statistical independence from the target variable (occupancy).
-
LASSO (Least Absolute Shrinkage and Selection Operator): A regularization technique that can perform feature selection by shrinking the coefficients of less important features to zero. The goal is to identify a minimal set of highly informative and uncorrelated features, reducing computational complexity and improving model interpretability.
-
Classifier: A simple Decision Tree classifier is used. The authors explore optimizing its parameters (like maximum depth) and also evaluate ensemble methods, ultimately settling on a simple, interpretable decision tree for its balance of performance and computational cost.
Datasets
The study utilizes a subset of a multi-sensor occupancy detection corpus collected at Robert Bosch GmbH. Specifically, they focus on:
- Single Office: Approximately 75 hours of audio data, representing a commercial environment.
- Living Room: Approximately 22 hours of audio data, representing a residential environment, collected in a laboratory setting.
The ground truth for occupancy (number of people) was obtained from synchronized video recordings.
Key Insights and Ideas
- Audio is Sufficient (for Presence): The research strongly suggests that audio signals alone can achieve very high accuracy for presence detection (up to 99%), making it a viable primary sensor for this task.
- Feature Importance: Contrary to common practice in speech processing, basic time and frequency domain audio features (like Energy, Brightness) and their statistical summaries (Mean, Max) appear more crucial for occupancy detection than MFCCs. This highlights the distinct nature of environmental sound analysis.
- Feature Selection is Key: Employing feature selection methods like Chi2 and LASSO significantly reduces the feature space, leading to improved accuracy and drastically reduced computational load without substantial performance loss. Selecting just a few key features can be highly effective.
- Privacy-Friendly and Cost-Effective: The reliance on audio makes the system inherently more privacy-preserving than camera-based solutions and leverages inexpensive sensors, aligning with the goal of affordable smart building technology.
- Segment Length Matters: The optimal time window for analyzing audio features for occupancy detection is longer than typically used for tasks like speech recognition, with 30-second segments showing significant improvements.
Key Results
- Presence Detection Accuracy: Achieved up to 96% for residential (living room) and 99% for commercial (single office) setups.
- Head Count Accuracy: Achieved 63% for the living room and 91% for the single office environment using an 11-dimensional feature set. With an optimized 30-second segment length, head count accuracy improved to 70% (living room) and 95% (single office).
- Feature Reduction: Using only 11 selected features (out of hundreds) resulted in satisfactory classification performance and significantly sped up training and testing times.
- Error Analysis: The study identified that errors often occur in scenarios with low activity, background noise (like TV), or when distinguishing between 0, 1, or 2 occupants, suggesting areas for future improvement.
Limitations and Future Work
- Head Count Accuracy: While presence detection is highly accurate, head count accuracy, especially in the living room, still has room for improvement.
- Environmental Noise: Background noise, particularly TV sound, is identified as a significant challenge. The authors suggest techniques to suppress TV noise or fuse audio with other sensors (like motion or temperature) in quiet or low-activity scenarios.
- Static Occupants: The paper doesn't explicitly detail how static occupants are handled, though the error analysis implies that lack of audio signal can lead to misclassification.
- Future Directions:
- Fuse audio with other sensors (motion, temperature) to improve head count accuracy and handle low-activity situations.
- Develop specific noise suppression techniques for common environmental sounds like TV.
- Explore more advanced feature extraction and selection methods.
- Investigate the system's scalability to more complex and diverse environments.
Summary updated