A few-shot learning methodology for improving safety in industrial scenarios through universal self-supervised visual features and dense optical flow

Abstract

Industrial safety aims to prevent and mitigate workplace accidents and property damage. One common approach to identifying and analyzing potentially risky situations involves the use of static cameras to capture images or videos of facilities and production processes. However, current state-of-the-art deep learning-based solutions require extensive labeled datasets and substantial computational power to detect these dangerous situations. To address these limitations, this paper presents DINOFSAFE, a methodology that combines dense optical flow and the DINOv2 model, a vision transformer that learns universal visual features without supervision. Our methodology demonstrates dual efficacy by both minimizing the manual labeling efforts necessary for model training and ensuring computational efficiency. Optical flow estimates the apparent motion of objects in the input video streams, while the DINOv2 model generates high-dimensional universal representations capturing their visual properties. Using these representations, we train simple linear classifiers to identify moving objects and categorize them. This information aids in identifying and preventing hazardous conditions in industrial settings, such as pedestrians crossing paths with forklifts, forklifts approaching dangerous areas, loads falling from forklifts, and similar situations. We tested our solution on real videos sourced from industrial environments, resulting in promising outcomes. Furthermore, we compiled a comprehensive dataset consisting of approximately 6 500 images, which we have made publicly available for research and development purposes.