Journal of Computing in Civil Engineering

Construction projects are capital-intensive and risk-prone, which can lead to serious claims and disputes. Thus, early identification and intervention of potential risks in contracts play significant roles in preventing conflicts in advance. However, traditional approaches are mostly limited to the simple task of predicting fragmentary information (e.g., a type of risk) from contracts. This study aims to predict comprehensive information to determine risk-handling actions by simultaneously performing three classification tasks (i.e., risk identification, risk allocation, and risk response). Specifically, the proposed multitask model is designed to integrate shared layers extracting general features for all three tasks with task-specific layers extracting relevant features of each individual task. Thus, this approach allows learning both common and specific features within a single network. For performance evaluation, experiments were performed on a data set of 2,586 contractual clauses from 10 construction projects, in which performance was compared with single-task models not only on the entire data set but also on the smaller number of data. The results revealed that the proposed model exhibited higher performance (mean weighted F1 score of 0.90 and accuracy of 0.78) than single-task models; furthermore, shared layers may better recognize hidden patterns for each classification task with the smaller data set (e.g., 0.04 higher mean F1 score and 0.09 higher accuracy for 250 samples). Thus, the proposed model can successfully implement three tasks simultaneously. When such information (e.g., risk types, responsible parties, and corresponding response strategies) is available in an early contract review, contracting parties shall determine specific risk-handling actions for proactive risk assessment and management in construction contracts.

Small- and medium-sized construction sites are not efficiently managed because of the lack of budget and labor in safety management. In this paper, a real-time struck-by hazards detection system based on computer vision is proposed to automatically detect onsite hazards at small- and medium-sized construction sites by analyzing far-field surveillance videos. The proposed system consists of image processing technologies such as object detection, object tracking, image classification, and projective transformation, considering actual small- and medium-sized construction site conditions. Images obtained from small- and medium-sized construction sites have a fixed scene, far-field conditions, and crowded characteristics. In the object detection, an object class suitable for far-field conditions was defined, and object tracking and class changes were applied to ensure field applicability. In addition, to apply projective transformation using fixed scene images, the representative point of the detected object was established. Moreover, for the real-time application of the entire system, appropriate models for each function were selected, and an optimized application process was presented. Consequently, the integrated system, which simultaneously performs hardhat-wearing detection, heavy-equipment operation detection, signal-worker arrangement detection, and heavy-equipment proximity detection, was developed. The performance of the proposed model was evaluated individually, and the feasibility of the proposed system was verified by attaching the qualitative results of the field application to the real small- and medium-sized construction sites. In the quantitative results of the hardhat-wearing detection part, an accuracy of 91% and the system robustness change according to the parameters were presented.

Recent research has focused on visualizing and analyzing massive visual data (images/videos) captured at construction sites to improve coordination, communication, and planning. One key approach is to generate reality models (point clouds) from ordered visual data using three-dimensional (3D) reconstruction pipelines and then compare them with the as-planned four-dimensional (4D) building information model (BIM). However, ordered photo collection requires a strict capture plan and happens only after a specific time interval. Additionally, the reality models demand a considerable amount of processing time. As a result, the construction project status is often unreported between the intervals. The random photos captured daily by construction practitioners from different parts of the projects are helpful in filling this void. Localizing them in the 3D reality model helps to extract valuable information on time for effective construction monitoring. This study develops a system that localizes random photos in reality models using computer vision and deep learning–based approaches. Specifically, given a set of photos and a point cloud pair, a deep learning network for a six-degrees-of-freedom (6DOF) camera pose regression is trained to take any random photo within the point cloud region and estimate its position and orientation. The network performance is enhanced through data augmentation by another generative adversarial network: the pix2pix generative adversarial network (GAN). Finally, the poses are refined through traditional vision methods such as Perspective-n-Point (PnP) pose computation with random sample consensus (RANSAC). The proposed method was evaluated on a construction site. The system could localize random images captured during construction engineers’ daily work with as low as 0.04 m position error and 0.70° orientation error. In the end, this paper indicates further applications of construction image localization in the context of progress, quality, and safety monitoring.

In numerous infrastructure health monitoring and inspection applications, swift and precise three-dimensional reconstruction of subsurface objects from ground penetrating radar (GPR) data is of critical importance, particularly given the recent advancements in perception modeling and the emergence of deep learning. Nonetheless, current research on the reconstruction of subsurface infrastructure scenes and objects faces limitations. Owing to the restrictions of conventional GPR data processing, these methodologies are prone to GPR data with noisy backgrounds and struggle to recreate noncylindrical objects. This paper investigates the back-projection (BP) approach for GPR-based three-dimensional (3D) subsurface target reconstruction and presents a learning model that formulates the reconstruction as an implicit BP from 2D to 3D representations, circumventing any preprocessing requirements in contrast to traditional techniques. The proposed learned model ultimately generates an explicit volumetric representation of the subsurface objects. Experimental results show at least a 33% enhancement in the performance of the proposed model compared to meticulously designed baselines.

Reliable and accurate information is crucial for decision making in construction projects. However, stakeholders driven by profit have the potential to manipulate information, compromising information authentication and integrity (IAI). Even worse, digitization in the construction industry has made IAI extremely volatile, e.g., by easily copying, modifying, and falsifying. This study aims to ensure IAI by proposing a four-layer blockchain-based framework for combining smart construction objects (SCOs)–enabled oracles and hash-based digital signature techniques to protect both on-chain and off-chain information. Four deployed smart contracts provide three assurance mechanisms, i.e., signature verification, public data validation, and SCO cross-validation, which have been tested to improve the tampering detection accuracy by 19%, 11%, and 27%, respectively. The contribution of this study is to illustrate how a reliable and flexible blockchain oracle system can be established with limited resources to handle the concomitant IAI problem and provide an in-depth understanding of the IAI assurance mechanisms from the proposed framework. Future research can be conducted to reduce the signature size, enhance scalability, and further increase detection accuracy.

The quality of prefabricated concrete (PC) components during the construction phase is crucial for project safety. However, manual inspections are no longer sufficient to meet the demands of efficient and large-scale quality inspections of PC components. While computer vision (CV) can quickly inspect the surface quality of PC components, it fails to effectively prioritize critical quality defects among different components. Treating all quality defects equally would result in resource wastage. To address the efficient detection of external quality in PC components during the construction phase, this study proposes an appearance quality diagnosis method based on object detection and multimodal fusion decision. By integrating human and machine intelligence in quality inspections and implementing multimodal fusion decision-making, the intelligent quality diagnosis method becomes more targeted. By utilizing image object detection, the accuracy of identifying quality defects reached 87.70%. The fusion decision approach combining human and machine intelligence is applied to make informed decisions regarding structures with quality defects. Through the utilization of point cloud data, high-precision quality inspections of problematic components with an accuracy of 0.1 mm have been achieved. The developed case library enables defect tracking and provides recommendations for optimization solutions. The results demonstrate that the proposed engineering quality diagnostic method can effectively and quickly identify quality defects in PC components and provide improvement suggestions.

Advancement in video technology has made it possible for instructors to provide students with applied knowledge of construction practice. While videos can stimulate students’ interest in the construction domain by providing opportunities to observe real-life construction work, videos can sometimes contain extraneous information that may distract learners from essential learning contents. Computer vision techniques can be utilized to detect and direct learners’ focus to important learning concepts in videos. This study investigated the design and usability evaluation of an annotated video-based learning environment designed to direct learners’ attention to significant learning contents. Faster R-CNN with VGG16 backbone was trained with 21,595 images to detect practice concepts within videos. A Visual Translational Embedding Network was trained with the object detector and 8,004 images to predict interactions between subjects and objects of the practice concepts. The object detection model could detect all instances of subjects and objects, making the model sufficient for interaction detection. Usability evaluation was conducted using questionnaires, verbal feedback, and eye-tracking data. Results of the usability evaluation revealed that cues, such as bounding boxes, texts, and color highlights, drew learners’ attention to the practice concepts. However, students allocated more attention to the signaled images than the texts. The study contributes to dual-coding theory and the cognitive theory of multimedia learning through the use of cues to select, organize, and direct learners’ attention to noteworthy information within videos. This study also provides insights into the key features of cues that can facilitate learning of construction practice concepts with videos.

A significant portion of the energy used in building operations is wasted due to faults and poor operation. Despite volumes of research, the real-world use of analytics applications utilizing the data available from building systems is limited. Mapping the data points from building systems to analytics applications outside the building systems and automation requires expert labor, and is often done in point-to-point integrations. This study proposes a novel method for using queryable information models to connect data points of building systems to a centralized analytics platform without requiring a particular modeling technology. The method is explained in detail through a software architecture and is further demonstrated by walking through an implementation of an example rule from a rule-based fault detection method for air handling units. In the demonstration, an air handling unit is modeled with two different approaches, and the example analytic is connected to both. The method is shown to support reusing analytic implementations between building systems modeled with different approaches, with limited assumptions of the information models.

Visual data comprising images and videos has become an integral aspect of construction management, potentially supplanting traditional paper-based site documentation. With the vast amount of image data generated in construction projects, an efficient retrieval system that not only enhances visual data documentation but also promotes reutilization is needed. Existing label-based image retrieval methods for construction images require manual labeling and ignore visual information. Moreover, other content-based methods that consider visual properties of construction images are limited to utilizing simple visual features of the entire image. This poses a challenge when attempting to retrieve target images from the same construction site or those involving similar construction activities, particularly considering that construction images often share similar visual properties. This research introduces a content-based image retrieval method that employs object detection to identify significant subregions within construction images and convolutional neural networks to extract refined visual features of these subregions. By simply inputting a query image, the proposed method can accurately retrieve target construction images of interest. The proposed method was validated through experiments designed to retrieve target images in both same-site and same-activity retrieval scenarios. The proposed method achieved the best mean average precision (86.4%). This technology could contribute to construction data management and decision-making processes by providing an efficient information retrieval system.