An Object Detection Approach for Unsafe Acts in Substations
Year: 2025
Type: conference-paper
Abstract: Power safety production is of paramount importance. Substation operations, as a critical component, require the deployment of intelligent video systems to supervise and manage potential violations during the production process. This necessitates the identification and detection of target objects within the surveillance videos. Addressing the challenge that the widely-used yet computationally expensive YOLOv5l model is difficult to deploy on edge computing devices, this paper proposes the Mblock-DDC-BiFPN method, which integrates lightweight design and multi-object detection capabilities. The method reduces the model’s computational cost by employing DepCBL and Mblock modules, while simultaneously enhancing detection accuracy through the introduction of the Efficient Channel Attention (ECA) mechanism and the Bidirectional Feature Pyramid Network (BiFPN) structure. Experimental results on a self-constructed substation object detection dataset show that the proposed Mblock-DDC-BiFPN method achieves a [email protected] of 98.0% and a [email protected]:0.95 of 76.5%. Moreover, the model’s parameter count and computational complexity are significantly reduced compared to the original YOLOv5l. The experiments validate the effectiveness and practicality of the proposed method in the field of power safety production, offering an efficient solution for edge device deployment.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7127039473
