Overhead object projector: OverProjNet

Abstract

Despite the availability of preventive and protective systems, accidents involving falling overhead objects, particularly load-bearing cranes, still occur and can lead to severe injuries or even fatalities. Therefore, it has become crucial to locate the projection of heavy overhead objects to alert those beneath and prevent such incidents. However, developing a generalized projection detector capable of handling various overhead objects with different sizes and shapes is a significant challenge. To tackle this challenge, we propose a novel approach called OverProjNet, which uses camera frames to visualize the overhead objects and the ground-level surface for projection detection. OverProjNet is designed to work with various overhead objects and cameras without any location or rotation constraints. To facilitate the design, development, and testing of OverProjNet, we provide two datasets: CraneIntenseye and OverheadSimIntenseye. CraneIntenseye comprises actual facility images, positional data of the overhead objects, and their corresponding predictions, while OverheadSimIntenseye contains simulation data with similar content but generated using our simulation tool. Overall, OverProjNet achieves high detection performance on both datasets. The proposed solution's source code and our novel simulation tool are available at https://github.com/intenseye/overhead_object_projector. For the dataset and model zoo, please send an email to the authors requesting access at https://drive.google.com/drive/folders/1to-5ND7xZaYojZs1aoahvu6BkLlYxRHP?usp=sharing.

Keywords

• Projection detection
• Deep learning
• Deep regression networks
• Camera modeling
• Image formation
• Overhead object