Practical passive visible light positioning with lightweight parameter updating

Passive visible light positioning (VLP) has garnered widespread attention thanks to its usability and ability to be device-free for users. However, existing methods lack generalization to new scenes in realistic applications varying in surroundings due to the substantial differences in data features. To this end, we propose a lightweight parameter update scheme inspired by knowledge distillation and contrastive learning. In particular, we initially train an original model with a convolutional neural network for the VLP system. Subsequently, leveraging sparse data from new scenes and knowledge gained from the original model, we train a new network and extract its encoder parameters. These parameters are then merged into the original model so as to adapt to new application scenes. To verify the effectiveness of the proposed scheme, we conduct extensive practical experiments, and the results show that it maintains a root mean square error (RMSE) of 22 cm under varying scenes on average, with the error reduced to 1/3 of that provided by the conventional methods.