Open Access Open Access  Restricted Access Subscription or Fee Access

A DNN-based Civil Structure Semantic Segmentation and Motion Information Identification Algorithm



Camera can capture large amount of visual information of target within its field of view theoretically. This capability allows a camera to act as a sensor for obtaining displacement or distortion of targets at a large range of distance. Considering that there are so many cameras in public society (including the camera in smartphone) and the cameras will record the vibration of structures in an earthquake event, which will be very helpful for structural seismic damage assessment post-earthquake. With the development of AI, people begin to use DNN (deep neural network) to refine more useful information from images and videos. Significant progress in deep learning began in 2012 as computer computing efficiency increased significantly. Since then, deep learning gradually replaced many traditional methods and has become a new direction for many algorithms, including computer vision algorithms. DNN is one of the earliest algorithms to play its role. It has been used widely in the fields of image classification, semantic segmentation, target tracking and so on. In this paper, a DNN is designed to identify the displacement field of a frame structure model subjected to earthquake ground motion on a shaking table. An efficient training sample generation method is also proposed to help train the network. The network is based on traditional network U-Net, which can tell the semantic segmentation information from images. The network consists of an hourglass-type structure with loose sides in the middle. This structure provides sufficient feature extraction capability for the network and allows it to output an image of the same size as the input, which is semantic segmentation information. Another branch of the network focuses on comparing the motional connections between two adjacent frames of video. The displacement field as well as semantic segmentation results of the frame structure model is finally obtained through this network


Full Text: