To address the issue of goods position estimation in medium and large intelligent warehouse systems, this paper proposes an indoor visible light positioning model based on contrastive learning and received signal strength (RSS), namely the contrastive learning transformer (CLTf) model. First, hundreds of optical power values are filtered to select the highest-intensity light-emitting diode (LED) signals for constructing the optical power vector. Then, the Transformer model is employed to capture long-sequence dependencies, while contrastive learning techniques are integrated to mine anchor point prior knowledge for feature representation optimization. The simulation results show that in a medium to large warehouse space of 50 m×20 m×3 m, the average positioning errors of the CLTf model on the 1st to 5th shelves are 0.292, 0.344, 0.375, 1.133, and 2.471 cm, respectively, with a positioning accuracy of centimeter level, significantly better than traditional methods.