PDGV-DETR: Object Detection for Secure On-Site Weapon and Personnel Location Based on Dynamic Convolution and...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Public safety relies on accurate detection of weapons and personnel for risk assessment.
• Existing object detection models struggle with occlusion, small object detection, and complex backgrounds in security surveillance.
• These limitations lead to poor adaptability, low accuracy, and insufficient robustness in real-world security scenarios.

Purpose of the Study:

• To propose an optimized threat object detection framework (PDGV-DETR) for security surveillance.
• To enhance the detection and positioning of weapons and personnel in static images.
• To address challenges like occlusion, scale variation, and background interference.

Main Methods:

• Developed a dynamic hierarchical channel interaction convolution module to improve detection of occluded objects.
• Constructed an improved bidirectional hybrid feature pyramid network with cross-scale fusion for multi-scale feature expression.
• Introduced a global semantic weaving and elastic feature alignment network to enhance object-background discrimination.

Main Results:

• PDGV-DETR achieved a peak mAP50 of 85.9% on a conflict scene dataset, outperforming baseline models.
• Demonstrated statistically significant performance improvement over RT-DETR (p < 0.01) and Deformable DETR (15.1% accuracy increase).
• Achieved 93.0% mAP for gun and knife detection on the OD-WeaponDetection dataset, improving by 2.2% over RT-DETR.

Conclusions:

• PDGV-DETR effectively balances positioning accuracy, detection performance, and computational efficiency.
• The framework shows superior generalization and stability in complex security scenarios compared to general object detection models.
• PDGV-DETR provides a robust solution for object-level threat detection in security surveillance, supporting applications like public monitoring and risk warning.