Jiepeng Liu

This study develops a YOLO–VLM collaborative method for construction safety risk identification in quadruped robot-based inspection scenarios. The method combines high-frequency edge-side screening of visually explicit risks, asynchronous low-frequency key-frame recognition of more complex risks, and structured output recording within a mobile inspection workflow. YOLO is used to detect explicit targets, including helmets, high-visibility vests, fire, and smoke. The VLM performs domain-constrained recognition across seven predefined safety risk categories on sampled key frames. Experimental results show that YOLOv12n achieves an mAP50 of 0.865 and supports 30-fps video detection after edge-side deployment, while the VLM achieves a macro F1-score of 0.7969 across the seven predefined safety risk categories. A further comparison on a hazard-containing image subset shows that the collaborative setting improves overall recognition performance by strengthening the screening of visually explicit risks while retaining the VLM's capability for more complex risk recognition. A field-test example from a real construction-site inspection scenario further supports the applicability of the proposed method to mobile inspection tasks requiring both explicit-risk screening and more complex risk recognition.

With the development of intelligent construction, three-dimensional (3D) point cloud data (PCD) simplification has played an important role in back-end scenario analysis, reducing the computational and storage burden. However, the existing simplification methods are not applicable to the modeling analysis tasks in the architecture, engineering and construction industry. To address this issue, this study proposes a new simplification method, named as the frequency domain coding-based maximum difference simplification (FDCMDS). The FDCMDS is able to convert PCD into frequency domain multidimensional features to capture fine-grained structural variations of PCD. To improve the simplification efficiency, a 3D Canny key point detection combined with PCD gradient is proposed as a key point extraction module. Finally, a method for evaluating PCD density is designed by combining existing metrics. The validation experiments on PCD with different density distributions and volumes prove the effectiveness and feasibility of the proposed method.

2D-3D cross modal retrieval (CMR) aims to retrieve query image matching points from a 3D reference map. Existing classical CMR datasets and methods commonly support database-based retrieval only, i.e., the point cloud retrieval results are fixed-scale geometric surfaces. The failure to consider geometric regions and information scales fundamentally limits the practical deployment of CMR in engineering systems that require dynamic spatial reasoning, such as autonomous navigation or three-dimensional industrial measurement. In this article, we introduce a new benchmark called cross modal regional retrieval, which extends the classic CMR to allow the free retrieval of associated regions within the point cloud from images. Toward this, a multiview training paradigm is proposed in the training phase, which enables the model to identify occluded points in the region based on a single view. Autoencoders are utilized to learn the mapping of fusion features from a single view to multiple views. We also convert the image retrieval task within the scene cloud into a point classification task in the image to implement global free retrieval. The information fusion and guidance provided by the global point cloud enhances the capability of image cross-modal retrieval. To match the input patterns of the model, we propose a method for constructing datasets from three benchmark sources. Extensive experiments demonstrate that our method achieves state-of-the-art performance compared to existing methods for 2D-3D cross modal regional retrieval.

Bubble diagrams containing key features and information are used for generative design of floor plans. The lack of reliable methods for automatically generating bubble diagrams significantly affects the smoothness of layout generation systems. To improve the time-consuming and unstable acquisition process, a novel method based on graph neural networks (GNNs) is proposed to generate various residential bubble diagrams. First, a dual-branch graph neural network (DBGNN) is introduced to learn the feature patterns of heterogeneous links, including connectivity and adjacency relations. Then, decentralized node sampling (DNS) and centralized node sampling (CNS) are proposed to enhance the local feature learning of DBGNN. Subsequently, a variational graph autoencoder (VGAE) is used to learn the implicit distribution of topological patterns, enabling the model to generate diverse outputs. Experimental results show that the proposed model performs excellently in two link prediction tasks, achieving 92.39% ACC-Door and 78.84% ACC-Wall, while also generating 50 distinct bubble diagrams, validating the effectiveness of the proposed method and demonstrating its outstanding application value.

Fabrication drawings are essential for design evaluation, lean manufacturing, and quality detection of precast concrete (PC) components. Due to the complicated shape of PC components, the fabrication drawing needs to be customized to determine manufacturing dimensions and relevant assembly connections. However, the traditional manual drawing method is time-consuming, labor-intensive, and error-prone. This paper presents a BIM-based framework to automatically generate the readable drawing of PC components using building information modeling (BIM) and multi-agent reinforcement learning (MARL). Firstly, an automated generation method is developed to transform BIM model to view block. Secondly, a graph-based representation method is used to create the relationship between blocks, and a reward mechanism is established according to the drawing readability criterion. Subsequently, the block layout is modeled as a layout optimization problem, and the internal spacing and position of functional category blocks are regarded as agents. Finally, the agents collaborate and interact with the environment to find the optimal layout with the guidance of a reward mechanism. Two different algorithms are utilized to validate the efficiency of the proposed method (MADQN). The proposed framework is applied to PC stairs and a double-sided shear wall to demonstrate its practicability.

Detecting construction quality issues is crucial for an infrastructure project. In this study, a data-driven intelligent system is proposed to improve the efficiency of quality detection and management. First, an intelligent system framework including mobile devices, application software, and deep learning algorithms is developed. Second, an improve YOLOv5 (HA-YOLO) based on the hybrid attention is proposed to improve the inference speed and accuracy of defect detection. In addition, an apparent quality dataset from the construction site is analyzed and produced. Finally, the stability of the intelligent detection system is tested under various shooting conditions to demonstrate its effectiveness and robustness. The experimental results on public datasets (NEU-DET) show that HA-YOLO achieves 79.8 mAP with a maximum improvement of 13.8% over other series models. In our task, the HA-YOLO improves the detection accuracy by 2.7% and 3.8% in the comparison with YOLOv5x and YOLOv8 algorithms, respectively.

Composite hollow–core slab floor with reinforced concrete topping (i.e. CHFT) is relatively new, which can be applied to various long-span structures. However, these systems are typically lightweight and exhibit low damping, posing potential serviceability concerns related to human vibrations. This paper presents a comprehensive vibration test on a 9 m-long CHFT system. Natural frequencies, damping ratios, and mode shapes of the floor system were obtained through modal tests. The peak acceleration, root–mean–square acceleration, maximum transient vibration value, and perception factor of the floor under heel-drop excitation were obtained through the perceived vibration test and were checked against the available design codes and standards. Sensitivity studies using the finite element method were made to investigate the vibration performance of the CHFT system. Analytical formulas for the fundamental frequency and peak acceleration were derived, which are therefore suggested for practical use. Additionally, an approach for evaluating the vibration serviceability of CHFTs is described.

The registration of bridge point cloud data (PCD) is an important preprocessing step for tasks such as bridge modeling, deformation detection, and bridge health monitoring. However, most existing research on bridge PCD registration only focused on pairwise registration, and payed insufficient attention to multi-view registration. In addition, to recover the overlaps of unordered multiple scans and obtain the merging order, extensive pairwise matching and the creation of a fully connected graph of all scans are often required, resulting in low efficiency. To address these issues, this paper proposes a marker-free template-guided method to align multiple unordered bridge PCD to a global coordinate system. Firstly, by aligning each scan to a given registration template, the overlaps between all the scans are recovered. Secondly, a fully connected graph is created based on the overlaps and scanning locations, and then a graph-partition algorithm is utilized to construct the scan-blocks. Then, the coarse-to-fine registration is performed within each scan-block, and the transformation matrix of coarse registration is obtained using an intelligent optimization algorithm. Finally, global block-to-block registration is performed to align all scans to a unified coordinate reference system. We tested our framework on different bridge point cloud datasets, including a suspension bridge and a continuous rigid frame bridge, to evaluate its accuracy. Experimental results demonstrate that our method has high accuracy.