IDBO-stacking indoor fingerprint localization algorithm based on uncertainty estimation

Wi-Fi fingerprint-based indoor localization has garnered significant attention due to its low cost, easy deployment, and high accuracy. However, indoor signal fluctuations and existing method limitations—such as single algorithm usage, low accuracy, and poor generalization—lead to suboptimal localization precision and stability. To address these issues, this paper proposes an improved dung beetle optimizer (IDBO)–stacking indoor fingerprint localization algorithm based on uncertainty estimation. In the offline phase, a hybrid filtering method combining bilateral and Gaussian filtering is first applied to the received signal strength indication (RSSI) fingerprint data for denoising and edge feature preservation, improving the stability of the fingerprint database. Next, the combination of base learners with complementary performance is constructed. An index combining the Pearson coefficient and cosine similarity (the PC index) is designed to select candidate learners with good localization performance and significant diversity as base learners. Meanwhile, IDBO is introduced, which integrates circle chaotic mapping, osprey optimization, variable spiral searching, and Gaussian–Cauchy mutation to perform hyperparameter optimization for the learners. Finally, the training of the base learners and the meta-learner is completed and saved for subsequent online localization. In the online phase, the RSSI signal of the target device is collected in real time and input into the trained base learners. Using the proposed dynamic weight allocation method with uncertainty estimation, the fusion weights of each base learner are dynamically adjusted according to their performance on different data subsets. The final positioning coordinates are output by the meta-learner. Experimental results show that the proposed algorithm achieves an average error of 1.04 m, reducing error by 13.46% to 36.54% compared with other ensemble methods and single algorithms. Moreover, the cumulative distribution function curve converges faster, demonstrating superior positioning performance and strong noise robustness. Moreover, validation in different environments further demonstrates the algorithm's strong adaptability and generalization ability.