Multi-Scale Feature Pyramid Network: A Heavily Occluded Pedestrian Detection Network Based on ResNet

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing pedestrian detection algorithms struggle with heavily occluded targets, leading to reduced accuracy.
• Crowded environments pose significant challenges for feature extraction in pedestrian detection.

Purpose of the Study:

• To enhance the detection accuracy of pedestrians, particularly in scenarios with heavy occlusion.
• To develop a robust feature extraction method for occluded targets in crowded scenes.

Main Methods:

• Proposed a multi-scale feature pyramid network based on ResNet (MFPN).
• Introduced a double feature pyramid network integrated with ResNet (DFR) to improve semantic information and contours of occluded pedestrians.
• Incorporated repulsion loss of minimum (RLM) to refine bounding box predictions by distancing them from unrelated ground truths.

Main Results:

• Achieved 90.96% Average Precision (AP) on the CrowdHuman dataset, outperforming existing methods.
• Demonstrated a 5.16% AP gain compared to the FPN-ResNet50 baseline.
• The proposed MFPN method significantly boosted the performance of pedestrian detection systems.

Conclusions:

• The MFPN effectively enhances features of occluded pedestrians, improving detection accuracy in crowded environments.
• The novel DFR and RLM modules contribute to clearer feature separation and more precise bounding box localization.
• This approach represents a significant advancement in state-of-the-art pedestrian detection systems.