High-Accuracy and Real-Time Fingerprint-Based Continual Learning Localization System in Dynamic Environment

In dynamic environments, localization accuracy often deteriorates due to an outdated or invalid database. Traditional approaches typically use Transfer Learning (TL) to address this issue, but TL suffers from the problem of catastrophic forgetting. This paper proposes a fingerprint-based Continual Learning (CL) localization system designed to retain old data while enhancing the accuracy for new data. The system works by rehearsing parameters in the lower network layers and reducing the training rate in the upper layers. Simulations conducted with fused data show that the proposed approach improves accuracy by 16% for new data and 29% for old data compared to TL in smaller rooms. In larger rooms, it achieves a 14% improvement for new data and a 44% improvement for old data over TL. These results demonstrate that the proposed CL approach not only enhances localization accuracy but also effectively mitigates the issue of catastrophic forgetting.