Abstract: Unsafe behaviors of construction workers are the key factor leading to safety problems in the construction process of water conservancy project. Most construction sites adopt on-site safety inspection, wearable equipment, real-time monitoring, and other methods to identify unsafe behaviors, but such methods are time-consuming, laborious, expensive with low information level, and unfavorable to the timely discovery and early warning of dangerous behaviors. This paper presents a new method of unsafe behavior identification suitable for large scenes of water conservancy projects based on computer vision and deep learning. First, an improved method of YOLOv5 is developed to solve the problem of missing and wrong detection of small objects in construction workers and machinery in large scenes, and a multi-object object detection model of construction workers and machinery is constructed. Then, based on the target detection model, recognition methods are suggested for each of the routine unsafe behaviors, such as being close to the static danger area, dynamic construction machinery, and not wearing safety helmet, etc. Engineering application verifies that our identification method strengthens the means and intensity of construction site control and improves effectively the level of water conservancy engineering construction safety and intelligent control.

Key words: water conservancy project, construction worker, unsafe behavior identification, computer vision, deep learning