IoT powered RNN for improved human activity recognition with enhanced localization and classification.

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Human Activity Recognition (HAR) and localization are critical research areas driven by smart devices.
• Sensor data from smart devices often contains significant noise, necessitating robust system design.

Purpose of the Study:

• To develop a noise-impervious and efficient system for human activity recognition and localization.
• To leverage multiple algorithms for enhanced performance in HAR and localization tasks.

Main Methods:

• Signal denoising using a Chebyshev type-I filter, followed by windowing.
• Parallel feature extraction for activity and location, utilizing the Boruta algorithm for feature selection.
• Data optimization with Particle Swarm Optimization (PSO) and training parallel Recurrent Neural Networks (RNNs) for HAR and localization.

Main Results:

• The system demonstrated exceptional performance on the Extrasensory and Sussex Huawei Locomotion (SHL) datasets.
• Achieved accuracies of 89.25% and 90.50% for HAR on Extrasensory, and 95.75% for HAR on SHL.
• Achieved accuracies of 95.75% and 91.50% for localization on Extrasensory and SHL, respectively, outperforming state-of-the-art methods.

Conclusions:

• The proposed system effectively handles noisy sensor data for accurate human activity recognition and localization.
• The integrated approach of filtering, feature selection, optimization, and parallel RNNs provides a robust solution.
• The system's superior performance on benchmark datasets validates its efficacy and potential for real-world applications.