Real-Time Monitoring of Mobile Construction Resources Based on Multiple Object Tracking

Vision-based detection and tracking are crucial techniques for effective management of construction resources. However, existing methods face challenges in achieving the necessary requirements of high accuracy and robustness while maintaining real-time tracking speeds. In addition, current studies can track only a few types and a limited number of resources, resulting in restricted construction information. To achieve the real-time monitoring of various mobile resources on construction sites, this study proposes a novel approach for multiple object tracking. The proposed method integrates an improved YOLOv8 detector enhanced by a Path Aggregation Network and a tracker that utilizes a twice-association technique with refined components. Additionally, a dataset encompassing 13 categories of mobile construction resources (including workers and 12 types of machinery) was established to train the model. To validate the performance of our method, an experiment was conducted using 12 construction-site videos. The tracking results were then compared with the baseline method, utilizing a set of evaluation metrics, including the higher-order tracking accuracy (HOTA) metric, which has been newly applied in the construction domain. The results of the proposed method achieved an average precision and recall of 99.09% and 97.41%, respectively, while the multiple object tracking accuracy, identification F1 score, and HOTA reached 96.49%, 95.63%, and 90.20%, respectively. The proposed method possesses outstanding real-time tracking capabilities. Furthermore, the method exhibits remarkable robustness in addressing challenges such as varying illumination, occlusion, and small objects, while also exhibiting strong generalization across diverse construction scenarios. Hence, the proposed method demonstrates outstanding performance and holds potential for further on-site applications, ultimately enhancing construction safety and efficiency management.