Objective In the handling process of the automatic guided vehicle (AGV), accurate visual positioning of tray holes, particularly those that are small, distorted, and of low contrast, poses a significant challenge. Therefore, a visual positioning method for tray holes in AGVs based on improved YOLOv5s is proposed.Methods Combining the channel shuffling operation improvement of ShuffleNetV2 and the CBAM attention mechanism improvement, the basic YOLOv5s framework is improved to focus on the sub-pixel level boundary blurred hole position areas in the deformation key region. Based on the SIoU loss function, attention is paid to the micro hole positions, and the spatial three-dimensional coordinates of the tray hole positions in the camera coordinate system are calculated to obtain the transformation relationship from the camera coordinate system to the hole position area coordinate system. The improved YOLOv5s framework is utilized to output the three-dimensional coordinates of the tray hole positions in the AGV robotic arm coordinate system.Results The experimental results show that the experimental method can effectively capture sub-pixel level positioning accuracy boundaries, with an absolute error of less than 0.03 cm and a relative error of less than 0.83%. The F₁ score and mAP index are 95.2% and 94.8%, respectively. The number of floating-point operations, parameter count, and model volume are 4.8 G, 2.6 M and 4.28 MB, respectively.Conclusion The experimental method effectively solves the visual positioning problem of small, deformed, and low contrast holes in the tray holes that need to be positioned, and improves the efficiency of AGV tray handling.