Construction industry has been infamous for being the most complex and most dangerous industry among all the industries. Having the ability to obtain as much as information, including safety related information, about the construction site has long been desired as they will be very helpful in planning the progress and ensure safety on site. However, on-site in person inspection is still the best practise available today, which can be labour intensive and ineffective. Recent research has been looking into utilizing artificial intelligence (AI) and computer vision (CV) to help in monitoring the site condition. However, these deep learning approaches require large amount of labelled data or images for training, which the construction industry is lacking. Moreover, most studies have only pu...[ Read more ]

Construction industry has been infamous for being the most complex and most dangerous industry among all the industries. Having the ability to obtain as much as information, including safety related information, about the construction site has long been desired as they will be very helpful in planning the progress and ensure safety on site. However, on-site in person inspection is still the best practise available today, which can be labour intensive and ineffective. Recent research has been looking into utilizing artificial intelligence (AI) and computer vision (CV) to help in monitoring the site condition. However, these deep learning approaches require large amount of labelled data or images for training, which the construction industry is lacking. Moreover, most studies have only put their robust model to be tested on test cases similar to their training sample. Scene variation, such as changes in time, weather, which are very common changes in construction sites, has not been studied. Therefore, this study is proposing to adopt transfer learning techniques to mitigate the data intensive training and dealing with different changes in scenarios. This study also did a sensitivity analysis on target domain data added for transfer learning and developed 2 efficient training strategy with transfer learning technique. The first approach is based on the learning curve discovered from the experiment results. The curve can be used to estimate new images needed for achieving a designated accuracy. And this curve is suggesting 80% of model improvements happens in adding just 200 to 400 data points, which is around 20% – 40% of the original training dataset size. The second approach would also take the model performance on the source domain into consideration, and it is suggesting that by adding around 30% of the original training dataset size of new data should yield a more balanced model for wider application.