Depth-Assisted Industrial Safety Monitoring Reducing False Alarms in Forbidden-Zone Violation Detection Using YOLO and Monocular Depth Estimation

Industrial safety monitoring has gained significant importance in modern manufacturing environments, especially in facilities where workers interact closely with heavy machinery, rotating components, and hazardous operational zones. Traditional computer vision–based safety systems rely heavily on 2D RGB detection models such as YOLOv8 and similar architectures. While these models demonstrate high accuracy in detecting personnel and PPE compliance, they inherently lack spatial reasoning regarding the actual physical distance between workers and machines. This limitation frequently results in false alarms, primarily caused by visual overlap, perspective distortion, and occlusions in complex industrial scenes.To address these challenges, this study proposes a depth-enhanced industrial safety monitoring framework that integrates YOLOv8 RGB-based detection with monocular depth estimation produced by DepthAnything V2. The proposed system first detects workers and PPE items based on RGB images and then evaluates the spatial alignment between detected bounding boxes and pixel-wise depth information. Specifically, the system computes the depth difference between the worker’s foot region and the calibrated depth of the hazardous machine zone. If the 2D projection indicates a forbidden-zone intersection but the depth consistency is low, the system suppresses the warning, preventing a false alarm.Experiments conducted in an aluminum coil production environment demonstrate that the proposed depth-assisted logic reduces false-positive alerts by 34–57% while maintaining sensitivity to genuine unsafe events. The solution requires no special depth sensors, stereo cameras, or LiDAR systems, making it highly practical for retrofitting into existing factory camera infrastructures. The results confirm that monocular depth estimation is sufficiently stable for improving safety decision-making in real-time industrial applications.