Coupled Network for Robust Pedestrian Detection With Gated Multi-Layer Feature Extraction and Deformable Occlusion...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks have advanced pedestrian detection.
• Detecting small-scaled and occluded pedestrians remains a significant challenge.

Purpose of the Study:

• To propose a novel couple-network for simultaneous detection of small and occluded pedestrians.
• To enhance pedestrian detection robustness against scale variations and occlusion.

Main Methods:

• A gated multi-layer feature extraction sub-network adaptively generates discriminative features for scale variation.
• A deformable regional region of interest (RoI)-pooling sub-network addresses pedestrian occlusion.
• Investigated channel-wise and spatio-wise gate units to improve feature representation.

Main Results:

• Ablation studies confirmed the effectiveness of both sub-networks.
• The coupled framework achieved promising results on pedestrian detection datasets.
• Achieved the lowest missing rates for small (40.78%) and occluded (34.60%) pedestrians on the CityPersons dataset.

Conclusions:

• The proposed couple-network effectively addresses the challenges of detecting small and occluded pedestrians.
• The method demonstrates superior performance compared to existing approaches, particularly in complex scenarios.
• The approach offers a robust solution for advanced pedestrian detection systems.