Optimizing Bounding Box Regression in Complex Backgrounds
Jun Chen, Haiyan Zhang, Anjun Yu, Yiwei Wang · 2024
In the context of electrical power operation sites, the automated and accurate detection of whether workers are properly equipped with safety gear, such as safety clothing, helmets, and safety ropes for high-altitude operations, is crucial for ensuring the safety of personnel. Traditional object detection models often struggle to accurately identify these critical pieces of safety equipment in complex environments, due to variable object sizes, overlapping objects, and substantial background noise. To specifically address these challenges, we propose a novel loss function-Enhanced Alignment Intersection over Union (EAIoU) loss. This loss function is designed to enhance the model's ability to discriminate between closely spaced and overlapping safety gear in real-time, significantly improving robustness against scale variations and complex backgrounds. In experiments based on the YOLOv8s architecture, our model achieved a mean Average Precision (mAP) of 92.1 % on a dataset specific to electrical power operations, markedly outperforming traditional IoU loss functions. Additionally, EAIoU demonstrated improved performance on the COCO dataset, indicating its generalization capability across various complex scenarios. This research not only enhances the accuracy and real-time performance of safety equipment detection at electric power operation sites but also opens new avenues for advanced object detection technology in complex, multi-object environments.