TSE-CNN: A Two-Stage End-to-End CNN for Human Activity Recognition

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Wearable sensors offer robust human activity recognition (HAR) for healthcare applications like elderly monitoring.
• Existing HAR methods struggle with accuracy for complex activities (e.g., stair climbing) and limited subject-specific data.
• Current neural network approaches for HAR often incur high computational complexity and power consumption.

Purpose of the Study:

• To develop an improved human activity recognition method addressing accuracy limitations and computational demands.
• To enhance recognition accuracy for challenging activities and scenarios with sparse training data.
• To reduce the power consumption associated with wearable HAR systems.

Main Methods:

• Proposed a novel two-stage end-to-end convolutional neural network (CNN) architecture.
• Implemented a data augmentation technique to improve model robustness with limited training data.
• Evaluated the method against state-of-the-art approaches for HAR.

Main Results:

• Achieved significantly improved recognition accuracy compared to existing methods.
• Demonstrated enhanced performance for activities with complex patterns, such as ascending and descending stairs.
• Showcased reduced computational complexity and power consumption.

Conclusions:

• The proposed two-stage CNN with data augmentation offers a superior solution for wearable human activity recognition.
• This method effectively addresses key challenges in HAR, including accuracy for complex activities and limited data scenarios.
• The advancements contribute to more efficient and accurate wearable health monitoring systems.