Vision-based monitoring methods have been actively studied in the construction industry because they can be used to automatically generate information related to progress, productivity, and safety. Object detection is essentially used in such monitoring methods to infer jobsite context. However, as many classes of construction entities exist in a job site, large amounts of image data are required to train a detection algorithm to detect each class object in images. Although image data augmentation methods using 3D models were proposed, publicly available 3D models are limited to some construction object classes. Therefore, this study proposes a three-dimensional reconstruction method to generate the image data required for training object detectors. To use the generated synthetic images as training data, a histogram of oriented gradient (HOG) descriptor of a target object is obtained from these images. The descriptor is refined by a support vector machine to increase sensitivity to the target object in test images. The performance of the HOG-based object detector is evaluated using real images from ImageNet. The result shows that the proposed method can generate training data more effectively than existing manual data collection practices.
Bibliographical noteFunding Information:
The authors would like to thank Kinam Kim for helping with image data acquisition and the anonymous reviewers for their valuable comments that helped improve this paper. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP; Ministry of Science, ICT and Future Planning , No. 2011-0030841 ) and The Korea Agency for Infrastructure Technology Advancement (KAIA, No. 17CTAP-C133290-01 ).
© 2018 Elsevier B.V.
All Science Journal Classification (ASJC) codes
- Control and Systems Engineering
- Civil and Structural Engineering
- Building and Construction