In order to improve the performance of feature extraction and online location matching in location fingerprints database of underground WLAN personnel positioning system, an adaptive kernel principal component analysis (AKPCA) algorithm is proposed. The AKPCA algorithm combines the optimal AP selection algorithm with the kernel principal component analysis (KPCA) algorithm, which makes the calculation of eigenvalue have certain sub-region adaptability, and effectively solve the problem that the eigenvalue dimension solved by the maximum likelihood estimation method in the KPCA algorithm is too simple for the partitioned location fingerprints database. According to the coverage status of AP signals in the region, the optimal AP selection factor can be used to calculate the optimal intrinsic dimension after the construction of the location fingerprints database.In the results, AKPCA algorithm is better than KPCA algorithm in the calculation accuracy of eigenvalue of each sub-region, and the confidence probability reaches nearly 100% when the positioning error is 4 m, which is higher than 91.4% of KPCA algorithm. In the comparison of memory occupation in the positioning process, the average memory usage of AKPCA algorithm is 0.832 GB, which is better than 1.278 GB of KPCA algorithm and other fingerprint matching algorithms.AKPCA algorithm is superior to other feature extraction algorithms in positioning accuracy.It can effectively reduce the resource consumption in the online positioning process of the positioning system. In the future research, we will strive to further improve the positioning accuracy of divided underground tunnel.