A Framework of Combining Short-Term Spatial/Frequency Feature Extraction and Long-Term IndRNN for Activity...

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Smartphone sensor data is increasingly used for human activity recognition.
• Challenges include distinguishing similar activities and variations in device placement.

Purpose of the Study:

• To develop an accurate and robust human activity recognition framework using smartphone sensors.
• To address the complexities of long-range temporal dependencies and location variations.

Main Methods:

• A novel framework combining short-term spatial/frequency feature extraction with a long-term independently recurrent neural network (IndRNN).
• Incorporation of group-based location recognition to handle variations in smartphone carrying positions.
• Utilized the Sussex-Huawei Locomotion (SHL) dataset for evaluation.

Main Results:

• The proposed method achieved 80.72% accuracy on the SHL dataset.
• Outperformed existing single-model approaches.
• An earlier version secured a top award in the SHL Challenge 2020.

Conclusions:

• The combined short-term and long-term feature extraction framework effectively enhances human activity recognition accuracy.
• IndRNN is suitable for capturing long-term patterns in sensor data for activity recognition.
• Location recognition is crucial for improving the performance of smartphone-based activity recognition systems.