用于大视场监控视频的小目标检测网络sod-Yolo

Puntos clave

• The central aim is to address the challenges of small object detection in large field surveillance videos, focusing on sample scarcity, weak features, and inaccurate localization.
• Proposed a strategy combining semantic guidance and virtual simulation for low-cost, high-quality data generation.
• Designed a feature enhancement mechanism with video differential preprocessing and multi-scale feature fusion.
• Implemented a decoupled head structure with an innovative α-CIoU loss function to optimize bounding box regression accuracy.
• SOD-YOLO outperformed mainstream models in real-world scenarios, achieving the highest AP75 of 13.5% and AP50s of 53.9% at construction sites.
• Achieved an optimal AP of 42.6% and AP75 of 29.5% in highway service areas.
• In a challenging university campus scenario, SOD-YOLO improved AP, AP75, and AP50s by 4.1%, 2.5%, and 5.0% compared to the baseline model YOLOv7.