In recent years, visible light positioning (VLP) techniques have been gaining popularity in research. Among them, the scheme of using a camera as a receiver provides a low-cost, high-precision positioning capability and easy integration with existing multimedia devices and robots. However, the pose changes of the receiver can lead to image distortion and light displacement, significantly increasing positioning errors. Addressing these errors is crucial for enhancing the accuracy of VLP. Most current solutions rely on gyroscopes or Inertial Measurement Units (IMUs) for error optimization, but these approaches often add complexity and cost to the system. To overcome these limitations, we propose a 3D positioning algorithm based on an attention mechanism convolutional neural network (CNN) aimed at reducing the errors caused by angles. We designed experiments and comparisons within a rotation angle range of ±15 degrees. The results demonstrate that the average error for 2D positioning is within 5.74 cm and the height error is within 3.92 cm, and the average error for 3D positioning is within 6.8 cm. Among the four groups of experiments for 3D positioning, compared to the traditional algorithm, the improvements were 7.931 cm, 15.569 cm, 6.004 cm, and 16.506 cm. The experiments indicate that it can be applied to high-precision visible light positioning for single-light imaging.