OPS-YOLO: A Lightweight Foreign Objectors Detection Model for Overhead Power System

The performance of modern railway transit systems heavily relies on the overhead power system (OPS). However, OPSs built on the ground are vulnerable to foreign objects that can pose a threat to railway safety. In this paper, an improved lightweight detection method called OPS-YOLO, which is based on YOLO-v5, is proposed to detect foreign objects in OPSs. First, the Ghostnet-v2 is improved to compress the lightweight backbone model to slim down the original YOLO-v5 and reduce the complexity and computation. Then, a parameter-free attention mechanism SimAM is introduced to enhance the detection precision of OPS-YOLO. Furthermore, the Focal-EIoU is chosen as the loss function of the OPS-YOLO to improve object border localization precision while accelerating bounding box regression speed. To evaluate the performance of OPS-YOLO, we generate a customized dataset of OPS scenarios by using augmentation methods to expand the existing samples and improve training results. Finally, we extensive experimental tests on the dataset are conducted to verify the proposed method. Results show that the proposed OPS-YOLO improves the mAP by 0.5% compared to the classical YOLO-v5s network. Additionally, OPS-YOLO compresses the model parameters by 37.9% and reduces the model size by 37.5%. It is demonstrated that OPSYOLO can achieve a balance between computational efficiency and detection accuracy, making it suitable for automating OPS supervision and reducing reliance on human inspection.