Chapter 4: Automatic Detection of Positions and Shapes of Various Objects at Construction Sites from Digital Images Using Deep Learning
At construction sites, many pictures are taken for inspection and management of construction. Objects such as construction machinery, signages, signboard, construction workers, etc., are captured, but their existence and locations must be manually detected by humans. Thus, it is desirable to automate the object detection process in order to improve the efficiency. For detection of objects from digital images, machine learning with feature values of images has generally been employed. However, this method requires determination of features and contents, which is learned by humans, taking much labour and making it difficult to achieve satisfactory accuracy. On the other hand, deep learning can automatically determine these features and contents of various objects from digital images so that it can reduce labour and can increase accuracy compared to the conventional machine learning. Therefore, this research aims to automatically detect positions and shapes of objects from digital images by using deep learning. First, a dataset is created and Single Shot Multibox Detector (SSD) is employed for an object detection algorithm to detect positions. Next, re-detection of positions is performed by fine-tuning the weights of SSD. In addition, using detected object information can improve the efficiency of filing images. The shape of an object can be assumed using fully convolutional network (FCN). In this research, construction machinery, workers and signages were detected from digital images taken at construction sites by the proposed method. The proposed method has shown better performance compared to the conventional machine learning method. Finally, object position detection and shape detection are overlapped and the result shows the visually detailed object detection.