Cross-Domain Human Activity Recognition Using Low-Resolution Infrared Sensors.

Area of Science:

• Computer Vision
• Machine Learning
• Human Activity Recognition

Background:

• Low-resolution infrared sensors are cost-effective for indoor monitoring.
• Recognizing human activities across different datasets (cross-domain) remains challenging.
• Few-shot learning offers a solution for limited training data.

Purpose of the Study:

• To investigate the feasibility of cross-domain human activity recognition using low-resolution infrared sensors.
• To evaluate a novel prototype recurrent convolutional network (PRCN) with a few-shot learning strategy.
• To assess the model's performance on real-world indoor activity datasets.

Main Methods:

• A novel prototype recurrent convolutional network (PRCN) was developed.
• A few-shot learning strategy was employed for classification.
• The PRCN was tested on two independent datasets of indoor human activities.
• Three different networks were compared as feature extractors.
• Cross-domain evaluation was performed between the datasets.

Main Results:

• The PRCN demonstrated effectiveness in cross-domain human activity recognition.
• The model achieved good performance despite variations in training data diversity.
• The chosen feature extractors influenced the overall network performance.
• The approach is viable for classifying up to eleven daily activities.

Conclusions:

• The proposed PRCN with few-shot learning is a feasible solution for cross-domain activity recognition from low-resolution infrared data.
• The model's robustness to training data diversity highlights its practical applicability in real-world indoor environments.
• This research contributes to advancements in low-cost, intelligent monitoring systems.