In a diffuse optical channel，the visible light indoor positioning is affected by first-order reflection and noise，and thus the positioning error in boundary region is relatively larger than that in interior region. To solve this problem，a positioning algorithm of subregional visible light indoor based on multilayer Extreme Learning Machine (ELM) was proposed in this paper，and the effectiveness of the proposed algorithm was verified by simulation experiments. Firstly，the first layer ELM based on the entire experimental region was established to calculate the entire positioning error. Secondly，the second layer ELM based on the magnitude and distribution characteristics of positioning error was established，and the entire experimental region was divided into boundary subregion and interior subregion. Thirdly，the third layer ELM based on the extracted boundary subregion was established to calculate the boundary positioning error. Lastly，the entire error with updated boundary error was used to realize the positioning. The experimental results show that the entire average positioning error of the proposed algorithm is 2.79 cm. Compared with the Received Signal Strength（RSS） and Back Propagation（BP） neural networks，the average positioning error is reduced by 13 times and 55.36%，respectively. Compared with the single-layer ELM，the boundary average positioning error is reduced by 65.66%，the entire average positioning error is reduced by 23.77%. Experimental results indicate that the boundary positioning error of the proposed algorithm is obviously decreased，which means the proposed algorithm has higher positioning accuracy and robustness，and is suitable for various positioning applications.