Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices

arXiv:2409.00078 · eess.SP, cs.LG, cs.NI · Submitted 2026-08-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices".

Jane: The paper was written by Zhe Tang, Sihao Li, Zichen Huang, Guandong Yang, Kyeong Soo Kim et al. from Zhejiang University of Technology and Binjiang Institute of Artificial Intelligence, Zhejiang University of Technology and Suzhou Institute of Industrial Technology and Xi'an Jiaotong-Liverpool University and University of Liverpool and D5 Data Co., Ltd..

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone! Today we're looking at a paper that's got a mouthful of a title: "Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices." Jane, I'm going to need your help unpacking that one.

Jane: Happy to, Tom! Let's break it down piece by piece. Indoor localization is how your phone figures out where you are inside a building when GPS doesn't work. This paper is about doing that using Wi-Fi signals, which is called fingerprinting.

Tom: Right, and the "decentralized" part is the big twist here. Instead of sending all your location data to one big central server, you're running the localization model right on the IoT devices themselves—like a Raspberry Pi sitting on a wall.

Jane: Exactly. And the "Sparse Gaussian Process" is the mathematical engine doing the work. A Gaussian process is a way of making predictions that also tells you how confident it is in those predictions. The "sparse" part means it doesn't need to use all the data it has—it picks a few smart representative points instead.

Tom: And "Reduced-Dimensional Inputs" means they're also cutting down on the number of Wi-Fi access points they look at. So they're trimming the data in two different ways to make it fast enough to run on small devices.

Jane: The really exciting part is what this enables. Because the model is small and fast, you can retrain it constantly with fresh data. Indoor environments change—people move, doors open, furniture shifts—so having a model that can adapt in real time is huge.

Tom: And that's something a big centralized system just can't do easily. Retraining a huge neural network on a server takes time and money. This approach lets each floor of a building have its own little model that's always learning.

Jane: Plus there's a privacy angle. Your location data never leaves the local device, so it's much harder for someone to track you across an entire building complex.

Tom: I love that. It's not just about making the tech faster—it's about making it more private and more resilient. If one device fails, you only lose localization for that one floor, not the whole building.

Jane: And that's the core promise of this paper. It's a framework that could make indoor localization practical, private, and adaptable in ways the old centralized approach never could be.

Tom: So stick around, because next we're going to dig into the actual results and see how well this thing performs in practice.

Summary: Jane: So Tom, we've talked about what this paper is trying to do. Now let's look at what they actually found. The paper "Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices" has some pretty impressive numbers.

Tom: I was just looking at the tables. On a regular server, their SGP-RI model with half the training data gets a 2D error of about five point eight meters, compared to five point three two meters for the full Gaussian process. That's only a nine percent increase in error, but they cut the training time by more than half.

Jane: And when they moved to the Raspberry Pi, the difference got even bigger. The full GP took over ninety-six seconds to train. The SGP-RI model with fifty percent sparsity finished in about twenty-four seconds. That's a massive improvement for a device that's basically a credit-card-sized computer.

Tom: And here's the kicker—the accuracy stayed competitive. On the Pi, they got a 2D error of five point eight four meters with that fifty percent sparsity model. The full GP got five point four four meters. So you're giving up a little bit of accuracy to get a model that trains four times faster.

Jane: But the real magic happened when they simulated a dynamic environment. They split the test data into four time periods, and they let the SGP-RI model retrain as new data came in. The DNN and CNN models, which couldn't retrain, saw their error climb from about five point six meters to over six meters over time.

Tom: But the SGP-RI model stayed around five point four to five point eight meters the whole time. It was actually the best performer across all four periods. That's the whole point—being able to adapt to a changing environment beats having a slightly better static model.

Jane: And they tested this on the UJIIndoorLoc database too, which covers multiple buildings and floors. Their model got a three dee error of six point eight seven meters, which is competitive with some of the top entries from the two thousand fifteen EvAAL competition.

Tom: So the summary is: you can get almost the same accuracy as a full Gaussian process, but with a fraction of the training time and on hardware that costs less than a hundred dollars. That's a pretty compelling package.

Jane: And it's not just about the numbers. It's about what those numbers enable—real-time adaptation, privacy, and resilience. We'll get into the specific improvements they made to achieve this next.

Tom: Right, because they didn't just throw a sparse Gaussian process at the problem. They made some clever choices about which data to keep and which to throw away.

Improvements: Tom: So Jane, we've seen the results. Now let's talk about how they actually pulled it off. The paper "Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices" makes two key improvements over a standard Gaussian process.

Jane: Right, and both of them are about being smart about what data you use. The first improvement is in how they pick which Wi-Fi access points to include. They call it WAP-based feature selection.

Tom: And it's surprisingly simple. They sort the access points by how much their signal strength varies, then they compare neighboring ones. If two access points have almost identical signal patterns—meaning they're probably in the same place or very close—they keep the one with more variance and drop the other.

Jane: The idea is that if two access points always give you the same reading, they're not giving you any new information. So you can safely get rid of one without hurting your accuracy.

Tom: And they have a clever way of doing this that's cheap to compute. They don't need to do any fancy optimization—just element-wise differences and a threshold check.

Jane: The second improvement is in how they pick the inducing points for the sparse Gaussian process. Instead of optimizing them with an iterative algorithm, which would be too slow for a Raspberry Pi, they just divide the floor into a grid and randomly pick a few points from each cell.

Tom: That's so simple it almost seems too good to work. But it does, because the grid ensures you get good spatial coverage. You're not accidentally picking all your points from one corner of the room.

Jane: And they showed that this works across different sparsity levels. Whether they keep fifty percent, forty percent, or thirty percent of the training data as inducing points, the accuracy stays in a reasonable range.

Tom: The other thing I like is that they're honest about the trade-offs. They tested different threshold values for their feature selection, and they showed that if you're too aggressive, you save time but lose accuracy. If you're too conservative, you get better accuracy but slower training.

Jane: So they picked a middle ground—a threshold of zero point eight five—that balances the two. It's a practical engineering decision, not just a theoretical one.

Tom: And that's what makes this paper stand out. It's not just proposing a new algorithm; it's showing how to make it work on real hardware with real constraints.

Jane: Next, we should look at the actual first page of the paper and see what the authors are really claiming. There might be some nuances we've missed.

Tom: Good idea. Let's dig into the details.

First Page: Jane: So Tom, let's go back to the very beginning of "Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices." The abstract and introduction lay out the problem pretty clearly.

Tom: And the problem is that conventional indoor localization relies on a centralized server. That server handles all the data, all the training, all the requests. And that creates three big issues.

Jane: First, it's expensive to update. If the environment changes, you have to rebuild the fingerprint database and retrain the model. That's a huge cost, so most systems just don't do it.

Tom: Second, it's a single point of failure. If the server goes down, nobody in the whole building complex can get location services.

Jane: And third, there's the privacy concern. A centralized server sees every location request from every user. That's a lot of sensitive information in one place.

Tom: The authors argue that IoT devices—like the Raspberry Pi—have enough computational power to run their own localization models. They're not as powerful as a server, but they're powerful enough for a smaller service area, like a single floor.

Jane: And that's the key insight. Instead of one big model covering a whole multistory building, you have many small models, each covering one floor. Each one can be trained and retrained independently.

Tom: They also mention that a lot of IoT devices are already deployed in the field. Access points running OpenWrt, Raspberry Pis, things like that. So the hardware is already there—you just need the software to take advantage of it.

Jane: And that's what they're providing. A framework that turns those existing devices into a distributed localization system.

Tom: One thing I found interesting in the introduction is how they frame the contribution. They say their unique contribution is the adaptability of the framework. It's not just about making the algorithm faster—it's about creating a system that can respond to change.

Jane: Right, because in a dynamic environment, the ability to retrain quickly is more valuable than having the best static accuracy. That's the whole thesis of the paper.

Tom: And they back it up with the experiments we talked about earlier. The SGP-RI model maintained its accuracy over time while the static models degraded.

Jane: So the first page sets up the problem, proposes the solution, and hints at the results. It's a solid foundation for the rest of the paper.

Tom: Let's wrap up our thoughts on this one and get ready for the next paper.

Conclusion: Tom: Alright, we've spent a good chunk of time on "Decentralized Indoor Localization Based on A Sparse Gaussian Process with Reduced-Dimensional Inputs for Real-Time Sensing and Training on IoT Devices." Let's pull it all together.

Jane: The big picture is that this paper shows you don't need a powerful central server to do good indoor localization. You can do it on small, cheap IoT devices, as long as you're smart about how you use your data.

Tom: And they were smart in two ways. They cut down the number of Wi-Fi access points they look at, and they cut down the number of training points they use. Both of those cuts made the model fast enough to train in real time on a Raspberry Pi.

Jane: And the accuracy held up. On the dynamic test, their model actually outperformed the static neural network models because it could adapt to changes in the environment.

Tom: The implications are pretty big. This could make indoor localization much more practical for places like hospitals, warehouses, and underground facilities where you can't rely on a central server.

Jane: And the privacy angle is important too. Keeping location data on local devices means it's not all sitting in one place where it could be stolen or misused.

Tom: Plus, the resilience aspect. If one device fails, you only lose coverage for that one floor, not the whole building.

Jane: The authors also mention some future directions. They want to improve the feature selection and inducing point selection methods, and they're thinking about a hybrid framework that uses both a central server and local IoT devices.

Tom: That hybrid idea is interesting. You could get the best of both worlds—the adaptability of local models and the global view of a central server.

Jane: For now, though, this paper makes a strong case that decentralized indoor localization is not just possible, but practical. It's a solid step forward for the field.

Tom: And with that, we're going to say goodbye to this paper and get ready for the next one. Thanks for listening, everyone!

Jane: See you next time!

Zhe Tang, Sihao Li, Zichen Huang, Guandong Yang, Kyeong Soo Kim, Jeremy S. Smith, Zhaowei Zhu, Qi Xuan

Zhejiang University of Technology · Binjiang Institute of Artificial Intelligence, Zhejiang University of Technology · Suzhou Institute of Industrial Technology · Xi'an Jiaotong-Liverpool University · University of Liverpool · D5 Data Co., Ltd.

eess.SP, cs.LG, cs.NI

Submitted: 2026-08-07

Comments: 9 pages, 4 figures, published in IEEE Sensors Journal

Journal ref: IEEE Sensors Journal, Early Access, Aug. 6, 2026

DOI: 10.1109/JSEN.2026.3718309

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 56/100

Key concepts

Indoor Localization
This is the process of figuring out a person's location inside a building when GPS signals are unavailable. The paper uses Wi-Fi signals, known as fingerprinting, to achieve this.
Decentralized
Instead of sending all location data to one central server, this method runs the localization model directly on the IoT devices themselves. This allows for local training and operation.
Sparse Gaussian Process (SGP-RI)
This is a mathematical engine used for predictions that also measures prediction confidence. The 'sparse' aspect means it uses only a few smart representative points instead of all the data, making it efficient.
Reduced-Dimensional Inputs
This refers to cutting down the number of Wi-Fi access points the system looks at. This reduces data input, making the model faster and suitable for running on small IoT devices.

Terminology

Summary

Summary

This paper proposes a decentralized indoor localization framework that leverages models based on a Sparse Gaussian Process with Reduced-dimensional Inputs (SGP-RI) deployed to Internet of Things (IoT) devices for smaller service areas, enabling quick adaptation to time-varying indoor electromagnetic environments through real-time sensing and retraining. The authors argue that conventional indoor localization based on a centralized server with substantial computational resources cannot easily adapt to time-varying indoor electromagnetic environments due to the high cost of fingerprint database updates and model retraining, and the centralized server is also susceptible to security breaches. The proposed framework addresses these issues by exploiting the redundant computational resources available on network and IoT devices (e.g., WAPs running OpenWrt, Raspberry Pi) for indoor localization.

The paper identifies key challenges in IoT-based indoor localization, including the need to strike the right balance between localization performance and resource utilization given limited computational resources and storage capacity. The authors note that data-sparse approaches like Bayesian models are more attractive than data-intensive approaches such as Neural Networks (NNs) because they can be trained with fewer training data for reasonable performance. They also discuss the issue of extrapolated RSSI values in fingerprint databases (e.g., -110 for undetected WAPs in the UJIIndoorLoc database), which may worsen model performance, particularly for Gaussian Process (GP) models that impose strict requirements on data quality and distribution.

The SGP-RI model is built upon GP regression formulated as a Bayesian linear model, which does not require backpropagation. Standard GP regression has cubic computational complexity O(N3) with respect to the number of training data N, dominated by the inversion of the covariance matrix. To address this scalability issue, the authors employ SGP, which incorporates a small number M (M≪N) of inducing points to reduce computational complexity to O(NM2). The paper introduces two key innovations: (1) WAP-based feature selection to reduce the dimensionality of input data (i.e., the number of WAPs W), and (2) RP-based inducing point selection to choose inducing points from the original training data.

For WAP-based feature selection, the authors propose a simple, heuristic scheme (Algorithm 1) that filters out WAPs not providing valuable features for localization. The algorithm sorts columns of the feature matrix based on column-wise variances in decreasing order, then compares adjacent columns by computing the element-wise absolute differences between their RSSI vectors. If more than 85% of the elements in the difference vector are less than or equal to 3, the algorithm removes one of the two columns based on which has the larger RSSI value at the index of the maximum difference. This reduces the number of columns from W to V, where the similarity threshold value of 3 and the comparison threshold value of 0.85 are determined based on experiments with two databases.

For RP-based inducing point selection, the authors propose dividing the covered area into a small rectangular grid and randomly selecting a certain number of inputs from each grid cell (Algorithm 2). This is designed for computationally-efficient implementation on resource-constrained IoT devices, avoiding iterative optimization algorithms like L-BFGS that are not suitable for such devices.

The decentralized indoor localization framework integrates the separate offline and online phases of conventional indoor localization into a unified workflow, providing continuous data collection and online instantaneous training. The floor-level database can be initiated with a limited number of samples from an existing database or constructed with newly-measured samples on-site at deployed IoT devices, then continually expanded and updated through crowdsourcing or by integrating unlabeled samples from users. The advantages of this framework are two-fold: it significantly reduces time and labor cost for database construction, and it enables location estimation to better reflect the time-varying nature of fingerprint statistics through continuous updates of both model and database.

The experimental evaluation uses two databases: the XJTLU dynamic database for single-building, single-floor indoor localization (covering three floors of the International Research Centre at XJTLU South Campus, with 101 RPs spaced about 3 m apart, measured over 44 days by surveyors and Raspberry Pi Pico Ws), and the UJIIndoorLoc database for multibuilding, multifloor indoor localization. The XJTLU database is split into a training dataset based on data from the first 24 days and a test dataset based on data from the remaining 20 days. The SGP-RI model uses the Rational Quadratic (RatQuad) kernel with scale mixture parameter α=2 and length-scale l=10.

Experiments on a server (AMD Ryzen 7 5800X, RTX 3060 Ti GPU, 16 GB RAM) show that the SGP-RI model with model sparsity of 50% reduces training time of the GP model by more than 50% (from 12.79 s to 6.08 s) at a slight increase in 2D error by about 9% (from 5.32 m to 5.80 m). The SGP-RI model with 30% sparsity further reduces training time to 5.00 s with a competitive 2D error of 6.44 m. Reference models include DNN, CNN, RF, k-NN, and Bonsai (a compact lightweight baseline for edge-oriented deployment). The Bonsai baseline reaches a 2D error of 7.28 m and requires 22.74 s for training at 80% model sparsity, indicating that its compact structure does not translate into a better accuracy-latency trade-off. RF and k-NN are computationally efficient but cannot provide decent localization performance.

Experiments on a Raspberry Pi 4B (Cortex-A72 CPU, 4 GB RAM, 16 GB storage) demonstrate that the SGP-RI model successfully completes training within 30 s and delivers reliable accuracy, despite the low computational capability. The GP model requires at least 64 GB of storage and active cooling to keep operating temperatures below 50 °C, conditions inconsistent with other models. The Bonsai baseline reaches a 2D error of 7.34 m and requires 120.47 s for training at 80% model sparsity, showing that SGP-RI is more suitable for frequent local retraining under the proposed decentralized framework.

Under a dynamic localization scenario, the authors simulated post-deployment conditions where the DNN and CNN models are trained on the first 24 days of data but tested on four groups of 5-day measurements each, while the SGP-RI, RF, and k-NN models sequentially move test groups into the training dataset to simulate frequent retraining. The results show that the SGP-RI model provides the best 2D errors over the whole period (5.46 m, 5.42 m, 5.64 m, 5.80 m for the four test periods), highlighting its capability to maintain higher localization performance despite environmental changes.

For multibuilding, multifloor indoor localization using the UJIIndoorLoc database, the SGP-RI model with 50% sparsity achieves a 3D error of 6.87 m, which is comparable to state-of-the-art models (e.g., RTLS@UM at 6.20 m, ICSL at 7.67 m, CDAELoc at 7.37 m, SALLoc at 8.28 m). The authors note that 3D errors should be interpreted as relative indicators because the top four models from the 2015 EvAAL/IPIN competition are evaluated based on the training, validation, and test datasets, while the rest are evaluated based only on training and validation datasets. Under the proposed decentralized framework, the building hit rate is set to 1 (assuming strong Wi-Fi signal attenuation between buildings), and the floor hit rate is estimated at 80% using a simple k-NN with k=7 for binary or ternary classification.

The sensitivity analysis of the WAP similarity comparison threshold in Algorithm 1 shows that lowering the threshold to 0.75 reduces training time (4.61 s) but increases 2D error (6.58 m), while raising it to 0.95 improves accuracy (6.11 m) at the cost of much longer training time (9.23 s). The threshold value of 0.85 is retained as a balanced default.

The paper also discusses case studies where the proposed decentralized indoor localization is particularly suitable: healthcare at hospitals/clinics (improved privacy for patients and service continuity), multitenant smart environments (physical separation of sensitive trajectory data into dedicated IoT nodes), dynamic industrial warehouses, disaster recovery, and underground mining (adaptability to layout changes, moving equipment, signal blockage, or emergency deployments).

The authors conclude that the proposed framework reduces reliance on a centralized server, which could be a single point of failure, and mitigates service disruptions caused by malicious attacks. Future work includes improving the WAP-based feature selection and RP-based inducing point selection, investigating a hybrid indoor localization framework utilizing both centralized server and multiple IoT devices, and addressing networkwide scalability, coordination between IoT nodes, and communication overhead for large-scale deployments.

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved system can do:

  1. Implement SGP-RI (Sparse Gaussian Process with Reduced-dimensional Inputs) as a lightweight regression module

  • Replace standard GP regression with SGP-RI using inducing points (30–50% of training data) and reduced input dimensions via WAP-based feature selection.

  • This reduces training time by >50% on servers and >75% on IoT devices while maintaining comparable accuracy (2D error increase of only 9%).

  1. Add a two-stage feature selection algorithm (Algorithm 1)

  • Filter WAPs based on variance, activity, and similarity (threshold 0.85, difference ≤3 dB).

  • This reduces input dimensionality from 466 WAPs to a smaller set, lowering computational cost without significant accuracy loss.

  1. Integrate RP-based inducing point selection (Algorithm 2)

  • Divide the service area into grid cells and randomly select a limited number of points per cell (e.g., 50% sparsity).

  • This avoids expensive iterative optimization (e.g., L-BFGS) and is suitable for resource-constrained devices.

  1. Enable real-time retraining for dynamic environments

  • Use a unified workflow where new fingerprint data is continuously added to the training set, and the SGP-RI model is retrained on-device.

  • This adapts to time-varying electromagnetic environments, as shown by the 2D error staying below 5.8 m over 20 days, outperforming static DNN/CNN models.

  1. Deploy a decentralized, multi-node localization framework

  • Run independent SGP-RI models on IoT devices (e.g., Raspberry Pi 4B) covering single floors or small areas.

  • This avoids single-point-of-failure risks and improves privacy by keeping location data local.

  1. Use the RatQuad kernel with fixed parameters (α=2, l=10)

  • This kernel provides stable performance across different floors and buildings without hyperparameter tuning, reducing deployment complexity.

  • Run on resource-constrained devices: Train a localization model on a Raspberry Pi 4B in under 30 seconds with 4 GB RAM, achieving 2D error of 5.8 m (vs. 5.44 m for full GP, which requires 64 GB storage and active cooling).

  • Adapt to changing environments: Retrain in real time as new Wi-Fi fingerprints are collected (e.g., hourly from fixed sensors or crowdsourced), maintaining accuracy over weeks without manual database updates.

  • Provide privacy-preserving localization: Keep all fingerprint data and model parameters on local IoT nodes, so user trajectories are not sent to a central server.

  • Handle large-scale multibuilding/multifloor scenarios: Achieve 3D error of 6.87 m on the UJIIndoorLoc database, comparable to top competition entries, using only 50% of training data.

  • Reduce operational costs: Eliminate the need for GPU servers and large storage, lowering infrastructure and energy costs for indoor localization services.

  • Support dynamic industrial/healthcare scenarios: Quickly adapt to layout changes, moving equipment, or emergency deployments without rebuilding a centralized database.

Sources

Related papers