HangCon: Benchmark Data Set for Enhanced Detection of Hanging Objects in Construction Sites

Abstract

Lifting operations on construction sites pose significant safety risks due to the potential hazard of falling objects. Effective monitoring of hanging objects is crucial for preventing accidents and ensuring worker safety. However, detecting hanging objects presents unique challenges for existing models, including the invariance in object shapes regardless of their hanging status, complex backgrounds that obscure ropes, and the diversity of hanging objects in terms of size, shape, and texture. To address these challenges, this study introduces HangCon (Hanging Objects in Construction Sites), a novel data set specifically designed for detecting “hanging objects”—loads suspended by tower cranes. HangCon contains 101,381 images, split between 50,842 images of hanging objects and 50,539 images of nonhanging objects, providing detailed annotations and diverse scenes. To evaluate HangCon’s effectiveness, this study conducted experiments using 10 benchmark models. The results highlighted the challenges in detecting hanging objects, with the best mAP at 71.63% for hanging objects alone, improving to 76.01% with unified annotations of objects and ropes. These findings highlight the complexity of detecting hanging objects and emphasize the necessity to implement advanced techniques such as semantic segmentation, depth estimation, and improved rope line detection. HangCon serves as a crucial resource for developing and refining detection models tailored to construction environments, significantly contributing to improved safety and operational efficiency on construction sites. By offering a comprehensive and well-annotated collection of images, HangCon facilitates the training and benchmarking of object detection models specifically for construction environments.