Two-Stage Pedestrian Detection Model Using a New Classification Head for Domain Generalization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning methods have significantly advanced pedestrian detection for applications like autonomous driving and surveillance.
• Existing detectors face challenges in domain generalization, limiting their performance across different environments.

Purpose of the Study:

• To develop a new two-stage pedestrian detector with improved domain generalization capabilities.
• To enhance feature representation by minimizing intra-class and maximizing inter-class distances using triplet loss.

Main Methods:

• Implemented a novel custom classification head with triplet loss integrated into Faster R-CNN and Cascade R-CNN architectures.
• Utilized the HRNet backbone pre-trained on ImageNet for feature extraction.
• Employed a progressive training pipeline, fine-tuning on progressively closer datasets to the target domain.

Main Results:

• Achieved state-of-the-art performance on the CityPersons benchmark.
• Obtained MR-2 scores of 9.9 (reasonable), 11.0 (small), and 36.2 (heavy).
• Demonstrated outstanding performance on the heavy subset, indicating robustness in challenging conditions.

Conclusions:

• The proposed triplet loss integration effectively improves domain generalization in pedestrian detection.
• The novel detector architecture and progressive training strategy yield superior results, especially in difficult scenarios.