| As one of the most representative manufacturing and real economy industries,the steel industry has played an important supporting role in the development of China’s economy.As the hub of iron and steel transfer in the production process,the warehouse is an important carrier of intelligent construction of iron and steel enterprises,in which the unmanned crane is the most important execution unit in the warehouse.At present,steel works have been relatively mature for the use of intelligent unmanned overhead cranes for steel coil handling,but in the process of steel coil handling,there may still be unmanned overhead cranes to lift the steel coil without clamping the steel coil so that the steel coil falls,resulting in major production safety accidents.Therefore,it is crucial to identify whether the crane clamp clamps the steel coil during the coil clamping process,which plays a crucial role in the safe operation of the unmanned crane.In the process of unmanned crane clamping the steel coil,the red area at the end of the clamp gradually disappears into the steel coil,and according to the object detection model,the process of tracking the disappearance of the red area at the end of the clamp can determine whether the clamp clamps the steel coil.One of the difficulties of this task is that the image captured by the unmanned crane in the night work scene is colorless,which is not conducive to the target detection model to locate the unmanned crane clamp and the red area at the end of the clamp.In order to solve this problem,this paper proposes to convert colorless images at night into colored images through image translation algorithms.The working range of the overhead crane is wide and the camera position is fixed,and the images taken by the overhead crane in different positions and under different light conditions may have that the red area at the end of the clamp is too small,which is not conducive to the detection of subsequent object detection models.Aiming at the problem of small detection difficulty at the end of the clamp in the image,this paper uses the Grad_CAM method to obtain the heat map of the working image of the unmanned crane,and on this basis,the area with high score is cropped to obtain an image containing only the clamp of the unmanned crane.Since the relative position of the unmanned crane clamp and the red area at the end of the clamp remain unchanged,the lower left and right corner areas of the cropped image are reparameterized to increase the weight of the red area at the end of the clamp and further improve the accuracy of the target detection model.In this paper,YOLOv5s is selected as the object detection model,and a large number of experiments are carried out on the real dataset of the full-scenario and full-time steel plant operation system to verify the effectiveness of the proposed method.The experimental results show that the image translation method proposed in this paper to convert colorless images into colored images improves the detection accuracy in night scenes from 42.23%to 66.3%;The improved object detection method in this paper improves the detection accuracy from 75.64%to 89.56%in daytime working scenarios and from 66.3%to 85.39%in nighttime working scenarios;After the above steps,the judgment accuracy of whether the unmanned crane clamp the steel coil is as high as 96.5%,and the detection rate reaches 5 frames per second,achieving the goal of high speed and high accuracy. |
