Game Engine-Based Synthetic Dataset Generation of Entities on Construction Site

Computer vision has been widely used in construction sites for progress monitoring, and safety monitoring. However, collecting data from construction sites and labeling them into datasets is a time-consuming, labor-intensive, and costly task. Therefore, a synthetic dataset generation approach for construction site entities based on the game engine is proposed to solve the problem of the lack of construction site datasets. In this research, construction site scene models are formulated by grouping existing digital on-site assets, and image annotation and camera calibration files are automatically generated by developed scripts in the selected game engine. The movement of the model is also controlled by developed scripts and the scene is rendered using High-Definition Rendering Pipeline (HDRP) to obtain high-resolution images. Components such as transform and Box Collider are used to get the coordinates of the object relative to the camera and the size of the bounding box, and to automatically generate the labels. In addition, the focal length, field of view (FOV), and other parameters of the camera component are utilized to calculate the camera Intrinsic when generating calibration files. By this method, a large amount of synthetic data can be quickly acquired and labeled, significantly reducing the time of dataset generation of on-site entities. Finally, the computer vision model trained on the synthetic dataset achieved 91.6% mAP on the real dataset.