Abstract: Replacing manual works with robots is a feasible solution for solving the safety problem of fastening bolts on the angle steel tower. In order to meet the operating requirements of the angle steel tower bolt fastening robot, a detecting and positioning system was proposed based on neural network and RGB-D camera for the main bolts of the angle steel tower. Applying the lightweight YOLOv5s-T network to the image of the Intel® RealSense™ depth camera D435i, the system was used to realize real-time detection, three-dimensional positioning and reordering the main bolts of the angle steel tower. Experiments show that YOLOv5s-T can improve the inference speed of the original algorithm by about 31% without reducing mAP (mean average precision) basically. Using three-dimensional coordinates measured by the RGB-D camera to calculate the distance between adjacent bolts, the average distance error is less than 1 mm. When the RGB-D camera is facing the bolt group template, the correct sorting rate of the template is above 95%. It can guide the end-effector of the 6-dof manipulator toward the target bolt within a short time.