Abstract: In order to identify potential accidents at construction scenes and process the pan-scene data on construction safety intelligently, the construction scenes were divided into two groups, unsafe behavior scenes and unsafe state scenes, on the basis of the scene factors. Using deep learning method, the object semantics, spatial relation semantics, scene semantics and behavior semantics were extracted from the pan-scene data on the basis of the division of image semantics hierarchies. Then, the correspondences were obtained between image semantics and the pan-scene data according to the theory of pan-scene data on construction safety. Building construction site photos were collected as the sample source to acquire the scene data of unsafe behavior in construction, and the feasibility of processing the pan-scene data by using a deep learning method was validated by analyzing the scene data of unsafe behavior in construction. The result shows that the pan-scene data can be acquired antomatically and processed efficiently through extracting the image semantics with a deep learning method.

Key words: deep learning, construction scene, image semantics, pan-scene data, unsafe behavior