DeepMonitor: A comprehensive approach for real‐time hazard detection for industrial safety

Abstract Recognizing workers' locations and automatically assigning warnings are crucial for preventing industrial injuries. However, existing warning systems are unsuitable for industrial environments as they rely solely on images or insufficiently leverage multi‐sensor inputs. Their distance‐ or plane‐based warning assignment strategies are limited when managing 3D spatial environments. To address these issues, we propose DeepMonitor, a novel industrial automatic warning system that incorporates a prior knowledge‐based 2D‐to‐3D conversion and a multi‐sensor, space‐based warning assignment strategy. We use a mature 2D object detector to avoid the need for 3D training datasets and apply prior knowledge with multi‐sensors to reduce the search space for workers' locations. To manage 3D spatial environments, warnings are assigned based on the overlap ratios between workers and zones, defined as 3D bounding boxes. We have constructed a novel dataset for industrial safety and have tested our system against existing approaches. Results demonstrate our system's superiority, achieving an F1‐score 16.7% and 24.7% higher than those of the image‐only and camera‐geometry systems, respectively.