Abstract: In order to improve the safety supervision of workers on the construction site, yolov5 target recognition algorithm combined with UAV tilt photography three-dimensional modeling technology are applied to create the intelligent detection model of unsafe state on the construction site to identify and position human, machinery and other targets. Through experimental comparison and analysis, the paper determines the optimal target recognition algorithm and constructs the multi-target recognition model. The results are in line with the theoretical conjecture, and the overall average recognition accuracy reaches 91.6%. After recognizing the target, its relative position is further determined with the spatial position information provided by the tilt photography three-dimensional model, which manages to ensure the safety state of workers. The accuracy of this visual positioning is determined by the three-dimensional model. It is verified that the distance error of the three-dimensional model constructed by UAV tilt photography is about 1.5% and the range length is greater than 35m, hence the distance error will be less than 1%, which notes the high accuracy for the distance of the object recognized by the target recognition model.