An Occlusion-Robust Feature Selection Framework in Pedestrian Detection ^†

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Advanced features enhance pedestrian detection but introduce noise and redundancy.
• High-dimensional features can hinder model learning and reduce performance.
• Existing methods often rely on dimension reduction techniques.

Purpose of the Study:

• To propose a novel feature selection framework for pedestrian detection.
• To enhance the handling of occluded pedestrians in detection models.
• To address noise and redundancy in high-dimensional feature sets.

Main Methods:

• Developed a simple yet effective feature selection framework.
• Introduced occluded pedestrian samples into the training dataset.
• Implemented a new feature selection criterion for occlusion handling.

Main Results:

• The proposed method demonstrated efficiency on the Caltech Pedestrian dataset.
• Achieved superior performance compared to state-of-the-art methods.
• Showcased significant improvements in detecting occluded pedestrians.

Conclusions:

• The feature selection framework effectively reduces noise and redundancy.
• The integration of occluded samples and a new criterion improves occlusion handling.
• The method offers a practical solution for robust pedestrian detection.