Source-linked AI summary
Low-effort place recognition with WiFi fingerprints using deep learning
Michał Nowicki, Jan Wietrzykowski
TL;DR
WiFi fingerprinting can localize indoors but traditionally requires dense surveys, filtering, analysis, and time-consuming tuning. The paper uses hierarchical deep neural networks with stacked autoencoders for building and floor classification, achieving comparable accuracy while reducing system-design effort on UJIIndoorLoc.
Problem
WiFi fingerprinting systems require dense radio maps, filtering, data analysis, and building-specific tuning, creating a substantial design burden.
Method
The paper applies hierarchical deep neural networks with stacked autoencoders that reduce raw WiFi feature dimensionality before building and floor classification.
Results
The best SAE-based architecture achieved 92% correct building and floor recognitions on testing samples, compared with 75% for a naive Euclidean-distance approach.
Takeaways & Limitations
Deep learning with stacked autoencoders produced comparable state-of-the-art accuracy while requiring no additional tuning or filtering.
Takeaways & Limitations
The evaluation uses UJIIndoorLoc measurements without averaging, lacks a dense radio map, and tests data gathered four months later.
Abstract
from arXiv · showhide
Using WiFi signals for indoor localization is the main localization modality of the existing personal indoor localization systems operating on mobile devices. WiFi fingerprinting is also used for mobile robots, as WiFi signals are usually available indoors and can provide rough initial position estimate or can be used together with other positioning systems. Currently, the best solutions rely on filtering, manual data analysis, and time-consuming parameter tuning to achieve reliable and accurate localization. In this work, we propose to use deep neural networks to significantly lower the work-force burden of the localization system design, while still achieving satisfactory results. Assuming the state-of-the-art hierarchical approach, we employ the DNN system for building/floor classification. We show that stacked autoencoders allow to efficiently reduce the feature space in order to achieve robust and precise classification. The proposed architecture is verified on the publicly available UJIIndoorLoc dataset and the results are compared with other solutions.
1 Introduction
Indoor localization lacks a universal solution, while common high-precision sensors can be expensive or computationally demanding. WiFi is widely available and can provide rough global position estimates, but fingerprinting requires surveying the environment and motivates lower-effort learning-based recognition.
- Laser scanners, cameras, and RGB-D sensors can provide precise localization but involve high cost or computationally demanding processing.
- WiFi adapters are common in mobile robots and personal devices, enabling rough global position estimates without additional exteroceptive sensors.
- WiFi localization requires surveying the environment to build a signal-strength map before operation.
- The paper investigates whether deep learning can recognize global location from sparse WiFi scans while reducing manual tuning effort.
2 Related work
WiFi fingerprinting supports indoor localization for robots and mobile devices, but established hierarchical systems depend on dense radio maps, filtering, analysis, and building-specific tuning. The paper proposes deep neural networks with stacked autoencoders to reduce feature dimensionality and workforce demands.
- WiFi localization has been used with particle filtering and odometry for mobile robots in buildings where other solutions are costly or unreliable.
- WiFi scans contain observed network identities, MAC addresses, and signal strengths, while ranging is hindered indoors by occlusions, reflections, and people.
- State-of-the-art fingerprinting commonly estimates building and then floor hierarchically, with building and floor usually predicted correctly in 85%−95% of cases.
- Precise kNN or weighted-kNN localization requires dense scans every 1−1.5 m, parameter searches, and time-consuming data filtering.
- Deep neural networks are proposed to reduce feature dimensionality, exploit larger datasets, and lower tuning effort for building and floor classification.
3 Low-effort place recognition with deep learning
The proposed system uses a hierarchical DNN that combines stacked autoencoders with a classifier for building and floor prediction from WiFi scans. Autoencoders learn reduced representations before classification, while dropout and softmax support generalization and class-probability outputs.
- Deep learning benefits from large training databases, and crowdsourcing makes large collections of WiFi scans feasible.
- The system follows a hierarchical approach and uses a DNN with autoencoders to predict building and floor from a single WiFi scan.
- A stacked autoencoder learns a reduced representation of raw WiFi measurements through unsupervised encoder-decoder training.
- After unsupervised training, the decoder is removed and fully connected classifier layers are attached to the encoder output.
- Dropout promotes redundant representations and generalization, while the final softmax layer outputs probabilities for the analyzed classes.
- Training uses separate training, validation, and testing sets with categorical cross-entropy and the Adam optimizer.
4 Experimental data
The evaluation uses UJIIndoorLoc, a large labeled WiFi dataset from the University of Jaume I, with training, validation, and testing partitions derived for comparison. Its real-life relevance is strengthened by device and user diversity, but delayed validation/testing and the lack of a dense radio map make localization challenging.
- Dataset composition: UJIIndoorLoc contains 21,048 WiFi scans, divided into 19,937 training and 1,111 validation samples.The scans cover almost 110,000 square meters across University of Jaume I buildings.
- Dataset composition: The dataset records measurements from 25 Android devices and 20 users.
- Input representation: Each scan has 529 attributes, including signal strengths for 520 discovered access points and nine measurement metadata fields.Signal strengths range from -104 dBm to almost 0 dBm; unavailable access points use 100.
- Evaluation split: The authors randomly split UJI training samples and treated UJI validation samples as testing data to enable comparison with competition results.
- Dataset conditions: Validation and testing occur four months after data gathering, and no dense radio map is available, making the dataset challenging but relevant to real-life performance.
5 Experimental evaluation
The experiments examine missing-signal representations, scaling, and DNN architectures for building and floor classification. Stacked autoencoders improve independent testing performance, with the best architecture reaching 92% correct recognitions after retraining on training and validation data.
- 5.1 Representing the lack of WiFi signal: The experiments represent missing access-point measurements as either 100 or -110 dBm and test alternative scaling procedures.
- 5.1 Representing the lack of WiFi signal: The best validation results use common scaling across WiFi measurements and represent a missing access point as -110 dBm.The authors attribute the advantage partly to avoiding discontinuity in the signal-strength representation.
- 5.2 Building and floor classification: Fully connected networks without autoencoders reached up to 98% correct validation recognitions but up to 88% on independent testing data.The authors associate the 10% gap with input vectors containing many missing-signal values.
- 5.2 Building and floor classification: Stacked autoencoders reduce the 520-network input to representations of 256, 128, and 64 features before classification.
- 5.2 Building and floor classification: The autoencoder architecture maintained around 99% validation recognition and increased testing recognition to as high as 91%.
- 5.2 Building and floor classification: 92% correct building and floor recognitions were obtained by SAE (256-128-64) + Classifier (128-128) after training on training and validation data.A naive Euclidean nearest-scan approach achieved 75%, while competition results ranged from around 87% to 96% on a different testing dataset.
6 Conclusions
The study applies deep learning to WiFi fingerprinting for building and floor estimation on a public dataset. Its conclusion is that stacked autoencoders support comparable accuracy with less tuning, while future work targets final-coordinate regression and finer within-floor localization.
- Conclusions: Deep learning estimates building and floor on a public dataset with accuracy comparable to state-of-the-art approaches and without additional tuning or filtering.
- Conclusions: Stacked autoencoders reduce WiFi feature dimensionality, producing networks that are easier to learn and perform better on testing data.
- Future work: Future work will estimate final X, Y position through regression and apply the architecture to smaller clusters or areas within each floor.