ARCOM

Search or browse again.

Click on the titles below to expand the information about each abstract.
Viewing 4 results ...

Croope, S V (2010) Managing critical civil infrastructure systems: Improving resilience to disasters, Unpublished PhD Thesis, , University of Delaware.

Dabash, M S (2022) Applications of computer vision to improve construction site safety and monitoring, Unpublished PhD Thesis, , University of Delaware.

• Type: Thesis
• Keywords: complexity; standards; monitoring; safety; supervision; training; construction site; inspection; regulation
• ISBN/ISSN:
• URL: https://www.proquest.com/docview/2778409265
• Abstract:
  This main objective of this research is to explore computer vision applications in improving construction site safety inspections and monitoring. More specifically, the research explores the current literature to identify limitations hindering computer vision application in the field and tackles three of the identified limitations pertaining to data availability, scene understanding, and data recording. The research identifies three limitations and attempts to answer three questions. Can super-resolution be used to enhance the quality of low-resolution progress images to train object detection models? Can we combine object detection models with ontologies to improve scene understanding and use the output of the models to perform safety inspections? And can we combine object detection models with facial recognition to create a database of unsafe behavior on construction sites?Computer vision capabilities in detecting objects on construction sites have been demonstrated by researchers working in this field. A significant limitation arises in scene understanding. Using object detection alone is insufficient to understand construction site safety behavior. This study presents a methodology combining computer vision and an ontology module to perform safety inspections. The methodology is demonstrated by developing a model for ladder safety inspections per OSHA and the American ladder association safe usage regulations. The output of two object detection models is used to infer the usage of ladders. An instance segmentation model identifies ladders in the scene and localizes them (Model Precision 97%, Recall 93%). A human pose model localizes the key points of the workers (Model Precision 96%, Recall 95%). Lastly, an ontologies model is fed the information from both models to assess the interaction and determine whether the interaction is safe based on regulatory standards (Model Precision 95%, Recall 98%).Implementing computer vision applications in the construction industry faces many obstacles, as identified by multiple researchers working in this field. A common obstacle is the lack of construction-related datasets and the general lack of data retention in the field. This study tackles data availability limitations in the construction industry by exploring the benefits of AI image enhancement techniques and their impact on computer vision detection models by training object detectors with varying quality datasets under the same condition to study the impact of image quality on the model's performance.A dataset of low-resolution ladder images was collected, sorted, and annotated. Subsequently, the low-quality images were run through different state-of-the-art AI image enhancement algorithms to raise their resolution. After preparing the datasets, they are used to train an object detection algorithm (YOLO v4) under the same conditions to study the impact of image enhancement on the model's performance. Image enhancement had a positive impact on the mAP. The D-DBPN enhanced model achieved the highest mAP at 90% and a recall of 90% relative to the low-resolution model that posted a mAP and Recall of 84% and 80%, respectively. The study also concluded that object complexity impacts the performance of SR image enhancement on object detection models.The construction industry is known for its high fatality rates relative to other less dynamic industries. 88% of construction site accidents are attributed to unsafe behaviors. Unsafe behaviors arise when workers do not follow safety rules, standards, procedures, instructions, and project-specific criteria; such behaviors can be limited or prevented with improved site supervision and management.This study proposes and tests a methodology combining computer vision-based object detection, tracking, and deep learning-based facial recognition systems to register the detected unsafe behavior in an SQL database. The proposed methodology is tested by developing a PPE detection model using YOLO v4 and combining it with a DeepSort tracking module and a facial recognition model to create database of workers performing unsafe behavior on construction sites. The detection model achieved a mAP of 93%, a recall of 91%, and the facial recognition module achieved a mAP of 99.38%. However, the performance highly depends on the detection angle.

Mayer, R H J (1982) Cost estimates from stochastic geometric programs, Unpublished PhD Thesis, , University of Delaware.

Soleimani, N (2022) Earthquake risk to civil infrastructure systems, Unpublished PhD Thesis, , University of Delaware.