Attention-Guided Sample-Based Feature Enhancement Network for Crowded Pedestrian Detection Using Vision Sensors.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occlusion, both inter-class and intra-class, is a significant challenge in computer vision-based pedestrian detection.
• Existing one-stage and two-stage detectors struggle with occlusion in complex urban environments, leading to reduced performance.

Purpose of the Study:

• To introduce a novel deep convolutional neural network architecture, the Attention-Guided Feature Enhancement Network (AGFEN).
• To address the limitations of current pedestrian detection methods caused by occlusion.

Main Methods:

• AGFEN enhances semantic information by mapping high-level features to low-level details via sampling, mimicking mask modulation.
• It improves both channel-level and spatial-level features without extra annotation costs.
• A one-to-multiple proposal-prediction paradigm and RoI feature aggregation using classification weights mitigate false positives.

Main Results:

• AGFEN demonstrated a 2.38% improvement over baseline detectors on the CrowdHuman dataset.
• Experimental evaluations on three benchmark datasets confirmed the effectiveness of the proposed architecture.

Conclusions:

• AGFEN offers a promising solution for improving pedestrian detection accuracy in occluded scenarios.
• The network's ability to enhance features and refine predictions holds potential for advancing computer vision technologies in real-world applications.