Aiming at the shortage of a single-photodiode (PD) receiver and geometric algorithms, we set up a real visible light positioning (VLP) scene of a multi-PD receiver and then use the fingerprint positioning technology based on the received signal strength, which commonly uses machine learning algorithms (MLAs). The positioning performance of four typical MLAs is studied. The results show that in two-dimensional positioning, the probabilities that the positioning error is less than 2 cm are 96.67%, 48.57%, 67.14%, and 15.24% for the K-nearest neighbor (KNN), extreme learning machine (ELM), random forest (RF), and adaptive boosting (AdaBoost), respectively, and in three-dimensional positioning, the probabilities that the positioning error is less than 2 cm for the KNN, ELM, RF, and AdaBoost are 74.52%, 38.81%, 59.76%, and 6.43%, respectively. Therefore, the positioning performance of the KNN is better in both the cases. On this basis, the influence of factors such as the number of light-emitting diodes (LEDs), number of PDs, and emission power of LEDs on the positioning accuracy is compared in detail. The results show that the increase in both the number of LEDs and PDs effectively reduces the positioning error. When the emission power of LEDs is 5 W, the positioning error convergence is achieved. The results provide a new theoretical support and practical application value for the design of VLP systems in the low LED distribution density scenes.