Pedestrian navigation activity recognition method based on two-stream transformer and contrastive learning.

Area of Science:

• Computer Science
• Robotics
• Signal Processing

Background:

• Pedestrian navigation activity recognition (PNAR) is crucial for pedestrian positioning and navigation.
• Existing methods face challenges in learning robust and generalizable sensor data representations.

Purpose of the Study:

• To propose a novel PNAR method combining a two-stream convolutional transformer with self-supervised contrastive pretraining.
• To enhance the robustness, transferability, and generalizability of PNAR models.

Main Methods:

• A two-stream architecture: spatial stream for multi-modal sensor dependencies and temporal stream for temporal relationships using attention.
• Self-supervised contrastive pretraining on unlabeled data to learn invariant representations.
• Evaluation on four public datasets and cross-dataset experiments.

Main Results:

• Achieved 99.08% accuracy and 99.22% F1-score, outperforming existing state-of-the-art methods.
• Demonstrated superior generalization ability in cross-dataset experiments with varying sensor configurations and activity labels.

Conclusions:

• The proposed two-stream convolutional transformer with self-supervised pretraining is effective for PNAR.
• The method significantly improves generalization, making it suitable for diverse real-world applications.