A crane drop zone safety warning framework based on an improved real-time detection transformer (RT-DETR) and stereo vision

Abstract To reduce the risk of workers entering the falling object hazard zone during crane operations and improve the efficiency of construction safety management, this paper proposes a three-dimensional detection and warning framework that combines an improved real-time detection transformer (RT-DETR) object detection algorithm with three-dimensional depth estimation. First, the RT-DETR object detection network is enhanced by incorporating the SSFF-P2 neck, the StarNet backbone, and the D 2 GCST feature interaction module. Compared with the original RT-DETR-R18, the proposed model raises mAP@0.5 by 1.0% and recall rate by 1.7%, while the parameters and floating-point operations per second (FLOPs) drop by 66.3% and 54.3%, respectively. These improvements strengthen the detection of small objects in real crane scenarios and simultaneously reduce computational overhead. In addition, a stereo-vision-based 3D distance estimation workflow is used to evaluate the risk of workers approaching the crane hook. By deriving disparity, the system reconstructs 3D coordinates and determines risk according to the measured distance between the worker and the hook. In the Monte Carlo uncertainty analysis, decision stability reaches 86.8% at a noise level of 0.5 m. The results demonstrate that our distance estimation predictions and safety decisions deliver reliable performance in real-world applications.