Abstract: In order to solve the problems of small detection range, complex detection process and susceptiblility to interference, a new detection method of construction tower crane is proposed by using DSOD(deeply supervised object detector)algorithm without model pre training. Firstly, a large number of clear tower crane images are collected from the construction site by camera to establish the data sets, and then the DSOD model is trained repeatedly to achieve the ideal convergence effect. Finally, the performance of this model is evaluated under different conditions(shooting angle, shooting distance and occlusion degree), and it is compared with the previous detection methods. The experimental results show that DSOD does have obvious advantages in detection accuracy and detection speed, and under the conditions of “front shooting”, “medium distance shooting” and “shielding degree less than 50%”, the precision and recall rate of tower crane detection can reach more than 90%, which provides a new idea for construction managers to solve practical problems of the project.

Key words: deep learning, object detection, construction hazard sources, tower crane, security management