Object Detection Based on Graph Neural Networks and Few-Shot Learning for Industrial Safety Control

Continuous monitoring of safety is possible with the use of modern video analytics systems. This allows monitoring the availability of personal protective equipment, monitoring the presence of people in hazardous areas, monitoring the safety of operations, monitoring compliance with work regulations. It is proposed to additionally use information about the basis of the scene structure using neural networks on graphs and few-shot learning methods to improve the performance of object detection. Each vertex represents a category in the knowledge graph, and the weight of the graph edge is calculated based on the conditional probability. The method contains two stages. The first stage performs object detection based on few-shot learning. Additional processing using a neural network on graphs is performed at the second stage. The method was studied on test images from the field of industrial safety control. Mean average precision increases by 1-5% when studying the developed method.