Inferring Human Activity Recognition with Ambient Sound on Wireless Sensor Nodes.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Wireless sensor networks (WSNs) with sound sensors offer context awareness in homes.
• Limited processing power in WSN nodes challenges signal feature extraction and sound source classification.
• Existing sound classification methods often overlook hardware constraints and algorithmic efficiency.

Purpose of the Study:

• To compare and evaluate various sound classification methods on a real sensor platform.
• To identify an efficient sound classification approach suitable for limited hardware.
• To assess the impact of feature types and classifiers on performance in resource-constrained environments.

Main Methods:

• Implementation and evaluation of multiple sound classification algorithms on a WSN platform.
• Utilizing diverse sound sources for realistic classifier training.
• Comparison of different feature extraction techniques and classification models.
• Analysis of algorithmic efficiency and performance under hardware limitations.

Main Results:

• Classifiers trained on limited hardware often exhibit lower quality.
• Sufficient classification performance is achievable by increasing classifier window length.
• A two-step classification approach (global followed by refined classification) enhances accuracy.
• The chosen methods demonstrate feasibility for real-time sound event detection in WSNs.

Conclusions:

• Optimizing window length and employing a two-step classification strategy are crucial for effective sound classification on limited WSN hardware.
• This research provides a practical framework for developing efficient sound-aware WSN applications.
• The findings contribute to advancing context-aware computing in smart home environments through optimized signal processing.