Automated tracking of worker and heavy equipment on tunnel construction sites: Deep-learning framework

Tunnel construction environments pose significant challenges for real-time monitoring due to narrow spaces, poor lighting, and high densities of personnel and machinery. This paper proposes an enhanced deep-learning framework for automated detection and tracking of construction workers and equipment in tunnel scenarios, aiming to improve safety and operational efficiency. We developed a specialized data set comprising 2,203 images and annotated ten categories of tunnel construction objects, including workers and various types of machinery. The original YOLOv8 object detection algorithm was optimized through three key enhancements: (1) self-calibrated illumination (SCI), which improves image quality in low-light conditions by enhancing visibility and contrast, thereby increasing detection accuracy under challenging lighting; (2) multiscale attention (MA), which enhances the detection of small and densely packed targets, a common issue in tunnel construction, by focusing attention on multiple scales simultaneously, resulting in better identification of workers and machinery in cluttered environments; and (3) the adoption of the lightweight GhostNet backbone, which reduces model complexity and improves detection speed without sacrificing accuracy. The improved YOLOv8 model was combined with the deep simple online and real-time tracking (SORT) tracking algorithm for real-time multiobject tracking in tunnel environments. Experimental results show that the improved YOLOv8 model achieved a mean average precision (mAP) at 0.5 IOU ([email protected]) of 0.907, outperforming the original model by 0.8%. Precision and Recall increased by 5.4% and 5.1%, respectively. Ablation studies confirmed the effectiveness of each enhancement: The MA module notably improved the F1 score by 3% and [email protected] by 2.2%. GhostNet reduced the model's parameters by approximately 43% and increased detection speed to 86.21 frames per second (FPS). The combined Deep SORT and improved YOLOv8 model yielded an overall ID F1 score of 70.6% and multiple object tracking accuracy (MOTA) of 71.5% across four testing videos, demonstrating robust tracking performance despite challenges such as occlusions and lighting variations. The proposed framework effectively addresses the complexities of tunnel construction monitoring, enhancing both detection accuracy and tracking reliability. This advancement enables real-time surveillance and analysis, supporting improved safety protocols and operational efficiency in tunnel construction projects.