Enhanced Abandoned Object Detection through Adaptive Dual-Background Modeling and SAO-YOLO Integration

Area of Science:

• Computer Vision
• Artificial Intelligence
• Public Safety Technology

Background:

• Existing abandoned object detection methods struggle with small and occluded objects, leading to high error rates.
• This impacts the effectiveness of public safety surveillance systems relying on automated detection.

Purpose of the Study:

• To develop an advanced abandoned object detection method to minimize false and missed detections, particularly for challenging small and occluded objects.
• To improve the overall accuracy and robustness of object detection in public safety applications.

Main Methods:

• An adaptive dual-background model was integrated with an improved Pixel-based Finite State Machine (PFSM) for enhanced background modeling and noise reduction.
• A novel Small Abandoned Object YOLO (SAO-YOLO) network was designed, featuring a Small Abandoned Object FPN (SAO-FPN) for comprehensive small object feature extraction and a Small Object Detection Head (SODHead) for precise local feature extraction and multi-scale fusion.

Main Results:

• SAO-YOLO demonstrated significant performance improvements, increasing mAP@0.5 by 9.0% and mAP@0.5:0.95 by 5.1% over the baseline model.
• Experimental results on ABODA, PETS2006, and AVSS2007 datasets showed an average detection precision of 91.1%, outperforming other advanced methods.
• The method notably reduced false and missed detections, especially for small and occluded objects.

Conclusions:

• The proposed SAO-YOLO method effectively addresses the limitations of existing systems in detecting small and occluded abandoned objects.
• The integration of the adaptive dual-background model and SAO-YOLO architecture significantly enhances detection accuracy and robustness for public safety surveillance.
• This approach offers a substantial advancement in automated threat detection, improving reliability in real-world scenarios.