TL;DR
This paper introduces a novel method for recognizing the daily activities of two residents in a smart home by leveraging time-based clustering and a similarity matching approach. By effectively de-noising sensor data and clustering activities based on their start and end times, the method achieves improved accuracy in distinguishing between individual activities, addressing a key challenge in multi-resident smart home monitoring.
What Problem Does This Paper Address?
The increasing global aging population presents a significant challenge for elderly care. Smart homes offer a promising solution by enabling activity recognition to monitor the well-being of elderly individuals, detect potential health issues, and provide timely assistance. While single-resident activity recognition has seen considerable progress, accurately distinguishing activities in multi-resident environments, especially when residents share spaces and activities overlap, remains a complex problem. This paper specifically tackles the challenge of recognizing activities for two residents simultaneously, aiming to overcome the interference and ambiguity that arises from multiple individuals' actions within the same smart home.
Methods
The proposed method consists of three main stages:
- Feature Extraction and De-noising: The raw sensor data is first processed to extract features. A key innovation here is a de-noising step designed to mitigate interference from the other resident's activities. This is achieved in two parts: first, by removing duplicate sensor features within a sequence, and second, by measuring sensor importance (using a metric called SIA) and retaining only those sensors that exceed a certain threshold, thus filtering out less relevant signals.
- Temporal Clustering: After feature extraction, the activities are clustered based on their start and end times. The K-means algorithm is employed for this purpose, with the optimal number of clusters (k) determined using the elbow method. This stage groups similar temporal patterns of activities together.
- Similarity Matching and Activity Recognition: For a given test activity, the method first identifies the cluster it belongs to. Then, it calculates a similarity ratio between the test activity and the activities within that cluster. This ratio is a weighted combination of temporal similarity (based on the difference in start and end times) and sequence similarity (using the Levenshtein distance on the sensor sequence). Finally, the activity is recognized by voting among the top 'n' most similar activities within the identified cluster.
Datasets
The authors evaluated their method on two publicly available multi-resident datasets from Washington State University's Center for Advanced Studies in Adaptive Systems (CASAS):
- Tulum2010: Collected from a smart apartment occupied by two married residents over 98 days, featuring 31 motion sensors and 5 temperature sensors, with 14 activity categories.
- Cairo: Collected from a voluntary adult couple's home (with occasional visits from children and a dog), featuring 27 motion sensors and 5 temperature sensors, with 13 activity categories.
Key Insights and Ideas
- Temporal Clustering is Key for Multi-Resident Scenarios: The paper highlights that focusing on the temporal aspects (start and end times) of activities is crucial for disentangling the actions of multiple residents, especially when they share common spaces.
- De-noising is Essential: Interference from other residents' activities can significantly degrade recognition accuracy. The proposed de-noising method, particularly the sensor importance-based filtering, is vital for improving the quality of the input features.
- Hybrid Similarity Metric: Combining temporal similarity with sequence similarity (Levenshtein distance) provides a more robust measure for matching activities, capturing both when an activity occurs and how it unfolds sensor-wise.
- Time-Based Clustering Outperforms Pure Sequence Matching: By clustering based on time, the method efficiently narrows down the search space for similarity matching, making it more effective than comparing a test activity against all training data.
Key Results
The proposed method demonstrated strong performance on both datasets, outperforming several common single classifiers and achieving competitive results with more complex methods.
- Tulum2010: Achieved an accuracy of 88.10%, precision of 88.00%, recall of 88.10%, and F-Measure of 87.90%. This was higher than Random Forests (RF), which was the next best performer at 83.60% accuracy.
- Cairo: Achieved an accuracy of 92.00%, precision of 92.90%, recall of 92.00%, and F-Measure of 92.00%. Again, this surpassed RF (89.60% accuracy).
- Qualitative Findings: The confusion matrices revealed that the method had very few errors for distinct activities. However, activities that shared similar sensor patterns and occurred in the same locations (e.g., "Enter Home" and "Leave Home" at the door) showed some confusion, which is a common challenge in activity recognition.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the cost of training might be higher due to the complexity of the method, though it claims it's better than many single and ensemble methods.
- Some activities with similar sensor patterns and temporal overlaps (e.g., "Eating" vs. "Enter Home") still present recognition challenges, as seen in the confusion matrices.
- The method is primarily tested on two-resident scenarios; its scalability to more residents is not explicitly discussed.
- Future Work:
- The authors suggest that further research could explore more advanced de-noising techniques and potentially more sophisticated clustering algorithms.
- Investigating methods to better differentiate activities that share spatial and temporal characteristics would be a valuable direction.
- Extending the method to handle a larger number of residents is a natural next step.
Summary updated