CIELO: Class-Incremental Continual Learning for Overcoming Catastrophic Forgetting With Smartphone-Based Indoor Localization

With dynamically evolving indoor environments, class-incremental learning (CIL) plays a crucial role in enabling indoor localization systems to adapt to new indoor areas. However, CIL poses additional challenges such as catastrophic forgetting, where patterns from previously learned paths are overwritten by data from new paths, and high data storage demands on the edge server, which must retain extensive localization data, resulting in high memory and power consumption overheads. To address these challenges, an effective solution must support CIL with indoor paths while mitigating catastrophic forgetting and reducing storage overheads on the edge server. To the best of our knowledge, CIELO is the first framework to address these challenges in the domain of indoor localization. It introduces a novel CIL approach that integrates an innovative representation memory management (RMM) policy with crowdsourcing to enable high-accuracy localization while significantly reducing catastrophic forgetting and data storage requirements. Through extensive experimental evaluations conducted across multiple real-world paths and devices, our results demonstrate that CIELO improves indoor localization accuracy by up to $29.4\times $ with up to 60 newly introduced classes (locations) across paths, reduces data storage by up to $1.75\times $ , and power consumption by up to $1.69\times $ on the edge server, compared to state-of-the-art solutions.