Session: 08-24-01: Human-Machine Interaction: Design, Dynamics, and Control

Paper Number: 167148

Human Detection and Localization in the Workplace Using a Single-Camera-Based Object Detection Algorithm 

Background
Accurate localization of human workers on factory floors enables better workflow monitoring and reduces hazards through timely detection of potential collisions with machinery. A camera-based system can track worker positions in real time, allowing management to optimize task allocations and respond proactively when unsafe conditions are predicted to arise. An image-based object detection algorithm such as YOLO X can be utilized for this application by enabling continuous monitoring of individuals and providing real-time location of each person.

Motivation and Purpose
Although many industrial applications rely on custom detection tools, a general-purpose, camera-based solution can offer flexibility in handling diverse factory floor layouts and worker movement patterns. By measuring how workers move and analyzing their motion paths, managers can streamline processes, detect bottlenecks, and predict hazardous scenarios such as collisions with robots or machinery. The main goal of this research is to apply the YOLO X object detection algorithm with a single camera for real-time worker tracking, enabling both workflow optimization and early warning of imminent collisions.

Contribution to Advancing Science and Engineering
While human detection and tracking utilizing object detection algorithms has been studied for numerous applications, accurate localization of individuals in 3D coordinates using a single camera system has not been thoroughly investigated. By utilizing a single camera with an open-source object detection algorithm, a cost-effective yet accurate human movement monitoring system can be realized. In addition, this paper presents a process for correcting localization errors that result from angled camera views, which can otherwise distort location estimates. This research demonstrates the potential for safer and more efficient industrial operations by combining worker detection with predictive modeling of worker movement.

Methodology
A camera is placed at an elevated location in a lab setting to capture continuous video of a simulated factory floor, where a subject person wearing markers performs designated movements. A marker-based 3D motion capture system tracks the person’s location and generates corresponding 3D coordinates. Meanwhile, the video images obtained by the single camera are processed by the YOLO X algorithm to generate bounding boxes and determine their centers in 2D image coordinates. Camera calibration is then conducted to convert these 2D detections into 3D coordinates in the workplace. Because angled perspectives can produce offsets between the bounding-box center and a worker’s true location, a regression-based technique is applied to refine the final positional estimates. The proposed single-camera system is evaluated under both static and dynamic conditions and the localization accuracy is measured by comparing the single-camera results with those from the marker-based 3D motion capture system.

Preliminary Results and Conclusions
Early tests show that the proposed system can detect a subject person reliably with minimal localization errors. By capturing each individual’s precise path, the system allows immediate responses when unsafe conditions arise, including predicting and preventing collisions between workers and mobile machinery including robot arms and forklifts. Moreover, the collected data enables managers to improve workflows through more informed decisions on task scheduling and resource allocation. Planned future work includes enhancing the collision prediction model and extending the system’s capabilities with additional data inputs with multiple moving subjects, enabling the way for a more comprehensive safety and productivity framework.

Presenting Author: Hwan-Sik Yoon The University of Alabama

Presenting Author Biography: Dr. Hwan-Sik Yoon is an Associate Professor in the Mechanical Engineering Department at The University of Alabama, Tuscaloosa, AL. His research interests include modeling, simulation, and control of automotive, transportation, and manufacturing systems. He is particularly interested in applying recent machine learning and reinforcement learning technologies to those systems. He is an Associate Editor of the SAE International Journal of Electrified Vehicles.

Authors:

Minjung Kim The University of Alabama
Shenglin Li The University of Alabama
Yun Chen The University of Alabama
Hwan-Sik Yoon The University of Alabama
Qiang Zhang The University of Alabama