Autonomous Visual Perception Framework for Threat Evaluation of Falling Object Hazard on Construction Sites

Author

• Nishita Sherathiya, University of Dayton

Date of Award

5-9-2026

Degree Name

M.S. in Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Vijayan Asari

Abstract

Falling objects from elevated heights remain a critical safety hazard on construction sites, contributing to hundreds of fatalities and injuries annually. While protective measures such as safety netting and toe boards provide passive mitigation, they often struggle to scale effectively in complex work environments. Furthermore, traditional safety monitoring relies on manual inspections that are labor-intensive, inconsistent, and incapable of providing real-time alerts. To address these challenges, this thesis proposes a vision-based framework for automatically detecting and assessing falling object hazards on construction sites using a monocular camera. The proposed system operates through a structured three-phase pipeline. In the first phase, a self-calibrating scene geometry module integrates monocular depth estimation (via DepthAnything V3) with semantic floor segmentation (SegFormer) and RANSAC-based ground plane fitting. This allows the system to recover real-world metric measurements without external calibration or specialized sensors.. In the second phase, construction tools are localized using YOLOv13 and precisely segmented using the Segment Anything Model (SAM). A specialized halo-based spatial analysis then estimates each object’s elevation and structural support percentage. Finally, a rule-based hazard classifier categorizes risks based on predefined height and stability thresholds. Experimental evaluation demonstrates that the proposed framework achieves a high degree of precision, with a Mean Absolute Error (MAE) of 3.471 cm and 2.5% in tool's height calculation and support assessment respectively. These results establish the system as a scalable, low-cost, and robust solution for enhancing occupational safety in high risk construction environments.

Keywords

Civil Engineering, Computer Engineering, Computer Science

Comments

OCLC No. 1591829454

Rights Statement

Copyright 2026, author.

Recommended Citation

Sherathiya, Nishita, "Autonomous Visual Perception Framework for Threat Evaluation of Falling Object Hazard on Construction Sites" (2026). Graduate Theses and Dissertations. 7693.
https://ecommons.udayton.edu/graduate_theses/7693