With the continuous development of artificial information technology, visible light positioning (VLP) based on deep learning has emerged as a hotspot for research on indoor localization technology. To improve the accuracy of indoor positioning, this paper proposes a convolutional neural network-gated recurrent unit (CNN-GRU) fusion neural network-based indoor VLP system. The system uses a single circular LED to transmit ID data and a common smartphone to capture feature images. To obtain light-emitting diode (LED) position information, this system adopts a method of extracting pixel values of the LED projection ellipse’s major axis for demodulation. In the coordinate prediction process, the feature parameters such as the location of the centre of mass, the length of the long and short axes, and the area of the ellipse in the image projection are input into the network for training, and the trained model is used to predict the coordinates of the test set. The results of this experiment show that the average positioning error is 3.79 cm in the 2.25 m × 2.25 m × 2.25 m experimental area, which can meet the positioning requirements of the indoor environment.