In order to solve the shortcomings of using the extreme learning machine (ELM) neural network to position indoor visible light, such as large error, long network model training time and poor stability of results, multi-objective momentum particle swarm optimization (MMPSO)-ELM scheme was formed by using sparse training fingerprint database, MMPSO and ELM indoor visible light positioning method. Momentum factor was introduced to avoid excessive oscillation during iteration and speed up the system convergence. Training data was set randomly in different positioning spaces. When the number of test points is different, the scheme of MMPSO-ELM was compared with back propagation, ELM and PSO-ELM. The simulation results show that, under the condition of 20 groups of training data and 80 points to be located, the maximum positioning error is 0.0225m, the minimum error is 0.00093mm and the average positioning error is as low as 0.00143m. The positioning performance is less affected by the size of the positioning space. MMPSO-ELM visible light positioning scheme has the advantages of high positioning accuracy, fast speed and strong generalization. This research provides theoretical support for fast and accurate positioning in indoor places.