Semester of Graduation

Summer 2025

Degree

Master of Science (MS)

Department

The Bert S. Turner Department of Construction Management

Document Type

Thesis

Abstract

Recent advancements in computer vision for construction site safety have encountered significant hurdles, especially in the nuanced tasks of object detection and the identification of unsafe worker behaviors. These challenges are often exacerbated by complex and cluttered backgrounds, wide variations in object scale, and inconsistent image quality. While existing methodologies have utilized attention mechanisms to analyze spatial and temporal features, they frequently neglect the benefits of adaptive sampling and channel-wise feature adjustments, thereby failing to exploit potential spatiotemporal redundancies. This thesis introduces a two-pronged approach to address these limitations. First, we propose the Optimized-Position Network (OP-Net), a novel architecture for object detection. The core of OP-Net is the Optimized Position (OP) module, which significantly enhances the relationships between feature channels by leveraging global feature affinity associations. This allows for a more robust and accurate detection of various objects on a construction site. Second, we present an innovative attention-based spatiotemporal sampling strategy designed to efficiently and accurately identify unsafe actions. This adaptive sampling method dynamically allocates computational resources, focusing on the most salient spatial regions and temporal moments in video data. This targeted approach minimizes redundant processing while maximizing the capture of critical information related to unsafe behaviors. To validate our proposed methods, we conducted extensive evaluations on two challenging, large-scale datasets. The object detection capabilities of OP-Net were rigorously benchmarked on the SODA (Site Object Detection) dataset, where it demonstrated superior performance in terms of both accuracy and efficiency. Furthermore, our unsafe action identification model was evaluated on the CMA (Construction Motion Analysis) dataset. The results show that our model not only achieves new state-of-the-art performance in accuracy but also maintains a reasonable level of computational efficiency, making it a practical solution for real-world deployment in construction safety monitoring systems.

Date

7-28-2025

Committee Chair

Wang, Chao

Download

Share

COinS