An Energy-Efficient Method for Human Activity Recognition with Segment-Level Change Detection and Deep Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Ubiquitous Computing

Background:

• Continuous human activity recognition (HAR) is crucial for context-aware services but is computationally and energy-intensive.
• Existing HAR methods with short analysis cycles struggle with the longer cycles of human activities, leading to inefficiency.

Purpose of the Study:

• To develop an energy-efficient HAR method with low computational complexity.
• To improve the efficiency of continuous activity classification in daily life.

Main Methods:

• Proposed a segment-level change detection technique to identify activity transitions efficiently.
• Utilized a fully convolutional network (FCN) for activity classification, triggered only upon detected activity changes.
• Compared the proposed FCN-based method against a convolutional neural network (CNN)-based approach using a public dataset on embedded platforms.

Main Results:

• The FCN model achieved a recognition rate comparable to the CNN model.
• The proposed FCN model requires only 10% of the network parameters compared to the CNN model.
• Energy consumption measurements showed the proposed method used up to 6.5 times less energy than the CNN-based method.

Conclusions:

• Segment-level change detection combined with FCN offers a computationally efficient and energy-saving solution for HAR.
• This approach significantly reduces the resource demands for continuous activity recognition on embedded systems.