Improving automation in computer vision based indoor construction progress monitoring: A deep learning approach
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Inefficient and inaccurate monitoring of work-in-progress significantly contributes to schedule and budget overruns in construction projects. Adoption of computer vision has expedited automation in monitoring construction progress by overcoming the challenges of laborious, costly and error prone manual methods. Majority of computer vision based progress monitoring studies have focused on the exterior construction environment and few studies have been conducted for indoor sites. The literature review found that robust object recognition from indoor site images is inhibited by technical challenges related to indoor objects, lighting conditions and camera positioning. Compared to conventional machine learning algorithms, deep learning models are robust. Therefore, this study aimed to develop a deep learning based approach for improving automation in computer vision based indoor construction progress monitoring.
Three indoor sites were used as case projects and time-lapse cameras were used to capture progress images. The framing, insulation and drywall installation of indoor walls was selected as the progress monitoring scenario. The images captured were first used to demonstrate the technical challenges and their impacts on feature extraction. Using transfer learning and hyperparameter fine-tuning, a Mask Recurrent Convolutional Neural Network (Mask R-CNN) model was optimised. Google Colab was used to train the deep learning model. Visual recognition of indoor as-built elements by the Mask R-CNN model was improved by optimising the learning rate with Adam optimiser and cosine annealing. Training efficiency improvements were made through the batch size, network resolution and number of iterations. The model reported a mean average precision of 88.02%, and training and validation losses of 0.19 and 0.20 respectively. This model was post-processed to automatically calculate the work-in-progress of indoor elements by extracting the areas of the segmented pixel masks.
The key contribution of this research is the development of an optimised and post-processed model of Mask R-CNN, which is capable of automated as-built visual recognition and work-in-progress measurement of indoor construction elements. This study also contributes to offering a simplified, efficient, and shareable training workflow on Colab that can be replicated in similar studies. The novelty of this research is in the optimisation process of the Mask R-CNN model and the Colab based training workflow. Upon providing training images, the Mask R-CNN model and the training workflow can be extended to other as-built scenes. The study provides explanations on the workflow of the deep learning model, which can be understood by construction professionals with a non-programming background.
Please use this identifier to cite or link to this item: