A Hybrid Framework for Predicting Crash Severity in Construction Work Zones Using Knowledge Distillation and Conditional GANs

Journal of Computing in Civil Engineering, Journal Year: 2025, Volume and Issue: 39(2)

Published: Jan. 13, 2025

Language: Английский

---

Deep Learning Image Captioning in Construction Management: A Feasibility Study

Journal of Construction Engineering and Management, Journal Year: 2022, Volume and Issue: 148(7)

Published: April 22, 2022

Deep learning image captioning methods are able to generate one or several natural sentences describe the contents of construction images. By deconstructing these sentences, object and activity information can be retrieved integrally for automated scene analysis. However, feasibility deep in remains unclear. To fill this gap, research investigates management. First, a linguistic schema annotating machine images was established, data set developed. Then, six from computer vision community were selected tested on set. In sentence-level evaluation, transformer-self-critical sequence training (Tsfm-SCST) method has obtained best performance among with bilingual evaluation (BLEU)-1 score 0.606, BLEU-2 0.506, BLEU-3 0.427, BLEU-4 0.349, metric translation explicit ordering (METEOR) 0.287, recall-oriented understudy gisting (ROUGE) 0.585, consensus-based description (CIDEr) 1.715, semantic propositional caption (SPICE) 0.422. element-level Tsfm-SCST achieved an average precision 91.1%, recall 83.3%, F1 86.6% recognition objects by generated sentences. This indicates that is feasible as generating accurate precise text descriptions images, potential applications analysis documentation.

Language: Английский

---

A universal traffic sign detection system using a novel self-training neural network modeling approach

Advanced Engineering Informatics, Journal Year: 2024, Volume and Issue: 62, P. 102674 - 102674

Published: July 3, 2024

Language: Английский

---

Tailored Vision-Language Framework for Automated Hazard Identification And Report Generation in Construction Sites

Published: Jan. 1, 2025

Language: Английский

---

Multi-Scale Fusion and Refinement Network for Precise Concealed Defect Detection

Published: Jan. 1, 2025

Language: Английский

---

Efficient semi-supervised surface crack segmentation with small datasets based on consistency regularisation and pseudo-labelling

Automation in Construction, Journal Year: 2023, Volume and Issue: 158, P. 105181 - 105181

Published: Nov. 18, 2023

Despite promising results in vision-based surface crack detection, data-driven approaches still suffer from the scarcity of rich labelled datasets. Such a limitation has hindered wider practical application detection models. To address this issue, semi-supervised framework is proposed, capable learning substantial amount unlabelled data and achieving high accuracy, even when available datasets are limited size. The designed by tailoring supervised training, consistency regularisation, self-training with certainty-based pseudo-labelling, resulting simple yet effective approach. using only 2% total Concrete Asphalt datasets, mIoU was 2.6% 4.7%, respectively, lower than best performance model trained on 100% data. Remarkably, assisted approaching exceeding saturation levels as little 20% 25%

Language: Английский

---

Deep learning-based object detection for dynamic construction site management

Automation in Construction, Journal Year: 2024, Volume and Issue: 165, P. 105494 - 105494

Published: June 15, 2024

Language: Английский

---

SRGAN-enhanced unsafe operation detection and classification of heavy construction machinery using cascade learning

Artificial Intelligence Review, Journal Year: 2024, Volume and Issue: 57(8)

Published: July 13, 2024

Abstract In the inherently hazardous construction industry, where injuries are frequent, unsafe operation of heavy machinery significantly contributes to injury and accident rates. To reduce these risks, this study introduces a novel framework for detecting classifying operations five types machinery. Utilizing cascade learning architecture, approach employs Super-Resolution Generative Adversarial Network (SRGAN), Real-Time Detection Transformers (RT-DETR), self-DIstillation with NO labels (DINOv2), Dilated Neighborhood Attention Transformer (DiNAT) models. The focuses on enhancing detection classification in through upscaling low-resolution surveillance footage creating detailed high-resolution inputs RT-DETR model. This enhancement, by leveraging temporal information, improves object accuracy. performance cascaded pipeline yielded an average first-level precision 96%, second-level accuracy 98.83%, third-level 98.25%, among other metrics. integration models presents well-rounded solution near-real-time dynamic environments, advancing technologies contributing safety management within industry.

Language: Английский

---

Construction Site Multi-Category Target Detection System Based on UAV Low-Altitude Remote Sensing

Remote Sensing, Journal Year: 2023, Volume and Issue: 15(6), P. 1560 - 1560

Published: March 13, 2023

On-site management of construction sites has always been a significant problem faced by the industry. With development UAVs, their use to monitor safety and progress will make more intelligent. This paper proposes multi-category target detection system based on UAV low-altitude remote sensing, aiming solve problems relying fixed-position cameras single category established targets when mainstream algorithms are applied supervision. The experimental results show that proposed method can accurately efficiently detect 15 types site targets. In terms performance, achieves highest accuracy in each compared other networks, with mean average precision (mAP) 82.48%. Additionally, applying it actual site, is confirmed have comprehensive capability robustness.

Language: Английский

---

Exploring Edge Computing for Sustainable CV-Based Worker Detection in Construction Site Monitoring: Performance and Feasibility Analysis

Buildings, Journal Year: 2024, Volume and Issue: 14(8), P. 2299 - 2299

Published: July 25, 2024

This research explores edge computing for construction site monitoring using computer vision (CV)-based worker detection methods. The feasibility of is validated by testing models (yolov5 and yolov8) on local computers three devices (Jetson Nano, Raspberry Pi 4B, Jetson Xavier NX). results show comparable mAP values all devices, with the processing frames six times faster than NX. study contributes proposing an solution to address data security, installation complexity, time delay issues in CV-based monitoring. approach also enhances sustainability mitigating potential risks associated loss, privacy breaches, network connectivity issues. Additionally, it illustrates practicality employing automated visual provides valuable information managers select appropriate device.

Language: Английский

---