Implementation of Artificial Intelligence for Detecting Modular Objects in Construction

Abstract

The practice of Artificial Intelligence (AI) is increasingly being promoted by technology developers. However, its adoption rate is still perceived to be low in the construction industry due to a lack of expertise and the limited reliable applications for AI technology. Both edge and object detection are crucial in construction for the achievement of the required information to significantly improve monitoring object installation progress as well as safety. There is a lack of studies on the monitoring of modular objects and evaluation of performance of algorithms for modular edge and object detection. Hence, this thesis aims to present detailed experimentation of the applicability and performance of edge and object detection algorithms on construction sites. To do so, this thesis provides a thorough evaluation of seven computer vision algorithms for edge detection and two machine learning algorithms for object detection on construction sites. To implement these experiments, four construction sites with different construction contexts are chosen to collect the required data for image analysis of the three types of objects including modular panels, safety barricades and construction fences. The construction sites include an educational building, a commercial building, and two segments of a light rail project in the northern and eastern suburbs of Sydney, Australia. A total of 1090 2D images and 21 3D models are collected using both 2D and 3D cameras for the imagery data acquisition. An advanced 3D camera is used to record details of the relevant spaces around the detected objects to help interpret the results, where required. The collected images are analysed by applying the edge and object detection algorithms. The edge detection algorithms including Sobel, Canny, LoG, Prewitt, Roberts, RGB-Sobel and fuzzy logic are fulfilled in MATLAB. The object detection algorithms include faster Region-based Convolutional Neural Network (faster RCNN) and Single Shot multibox Detector (SSD) from the open source machine learning library of TensorFlow in Jupyter environment. In order to evaluate the performance of these algorithms, several tests and metrics are used. Those for edge detection are: mean squared error; peak signal to noise ratio; structural similarity index; analysis of false positive and false negative errors; the degree of discontinuity; the ease of use for non-experts and the image processing duration metric. The tests used for object detection experiments include average recall and mean average precision by pixels, recall and precision by counting, and training loss. The evaluation in this thesis shows the best performer was the fuzzy logic method based on the proposed edge detection metrics. Also, it was found that the performance of faster RCNN and SSD depends on the context and object detection scenarios. Indeed, surrounding objects, backgrounds of the objects, lighting conditions and camera angles affect the level of accuracy obtained for the results. The major contributions of this thesis are: 1) to extend the applications of edge and object detection to analyse construction modular objects; 2) to investigate the performance of edge and object detection algorithms based on more selected metrics than previous studies; 3) to collect first-hand data using 2D and 3D cameras in construction field work. The collected data can be a valuable benchmark to construction scholars. The implication of this thesis is that the research can be used for developing further applications of AI technology in order to monitor safety on construction sites and physical progress of modular object installation, and to control installation quality as well. Further study is recommended to detect more types of construction objects so that they may be compared with the findings of this research.