Work

Intelligent system for illegal behavior safety supervision based on improved YoLov5 and generative visual big model

Year: 2025

Type: conference-paper

Abstract: There are a large number of violations of safety regulations at the construction site of the power grid. These behaviors do not comply with safety supervision norms and can cause significant safety hazards. In response to the problem of difficulty in capturing and alerting such violations, we have designed an intelligent recognition system for violations based on enhanced YOLOv5 and generative visual big models to automatically detect violations and potential safety hazards at the construction site of the power grid. We have constructed a private dataset of violations at the construction site of the power grid. In our improved YOLOv5 model, we first added a spatial channel attention mechanism to the Neck section of Yolov5 to enhance its ability to extract spatial features and distinguish between different channel features, enabling the network to better distinguish different targets. In addition, we have added a spatial attention mechanism after the SPP of the backbone network, which enables the network to fully utilize spatial information, extract more prominent features, and improve the accuracy and recall of the model. In addition, to compensate for the lack of negative samples in the dataset, we use a generative visual model to generate a large number of negative sample images to enhance the dataset and address the issue of insufficient negative samples. We have designed a real-time safety supervision intelligent system for detecting violations on construction sites based on the above model and method, using IoT devices such as on-site real-time cameras, intermediate data servers, and routers. We conducted extensive experiments on real construction site datasets, and the results showed that our method has significantly improved compared to the original model. The MAP of the system is 94.0%, indicating that the model has strong detection ability and robustness in identifying personnel and safety tools in power grid construction site scenarios.

Source: International Conference on Image Processing, Machine Learning and Pattern Recognition (IPMLPR)

DOI: 10.1145/3759928.3759934

Authors: Qiyou Lv, Zehong Chen, Xuliang Luo

Institution: Guangdong Power Grid Company (China)

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDGs: Industry, innovation and infrastructure, Sustainable cities and communities

Open Access status: bronze

Funder: National Natural Science Foundation of China