A Safety Detection Model for Substation Operations with Fused Contextual Information

Abstract

Detecting and regulating compliance at substation construction sites is critical to ensure the safety of workers. The complex backgrounds and diverse scenes of construction sites, as well as the variations in camera angles and distances, make the object detection models face low accuracy and missed detection problems. In addition, the high complexity of existing models creates an urgent need for effective parameter compression techniques to facilitate deployment at the edge server. To cope with these challenges, this study proposes a safety protection detection algorithm that fuses contextual information for substation operation sites, which enhances multi-scale feature learning through a two-path downsampling (TPD) module to effectively cope with changes in target scales. Meanwhile, the Global and Local Context Information extraction (GLCI) module is utilized to optimize the key information learning and reduce the background interference. Furthermore, the C3GhostNetV2 unit is utilized in discerning the interconnections of far-off spatial pixels, while enhancing the network's expressive power and reducing the number of parameters and computational costs. The outcomes of the experiments indicate that the present model improves upon the mAP50 metric by 4.5% compared to the baseline model, and the accuracy of the checks and the recall have seen respective increases of 4.8% and 10.1%, while there has been a reduction in both the count of parameters and the floating-point calculations by 16.5% and 12.6% respectively, which proves the validity and practicability of the method.