With the advent of smartphones and wearable devices, indoor positioning systems are relevant to the ageing in place and in-home monitoring paradigms. In terms of positioning, Wi-Fi is a widely known and used technology for this purpose. While fingerprinting in large areas benefits from a large wireless network with several Access Point (AP) providing network access, the availability of Wi-Fi signals with full or partial Line-of-Sight (LOS) conditions at home is mostly limited to a single AP. This work presents two datasets collected in an urban flat and in a home-like laboratory to support single contrast to other Wi-Fi datasets where the number of samples is limited in each reference point, the two datasets have several Received Signal Strengh (RSS) readings per reference and evaluation point in order to test advanced positioning methods based on Single-AP (S-AP) Wi-Fi fingerprinting.
Spain (V. K. Chaurasiya)
While several datasets exist for RSS-based fingerprinting indoors [
The main contributions of this paper include: • Full description of two new datasets for S-AP Wi-Fi fingerprinting; • Definition of baseline methods for both datasets;
The remaining of this work is organised as follows. Section 2 described the procedures to collect the datasets as well as the data format. Section 3 provides some examples of usage of the dataset, giving the baseline method for this dataset. Discussion and conclusions are provided in Section 4.
Data were collected in two environments: the urban flat (FLAT) and a research laboratory (LAB) environment. Both scenarios are partially seen in Fig. 1 (a) and (b).
(a) (b) (c)
Data were collected with an ESP32 board (see Fig. 1). In this case, the ESP32 development board was configured in station mode to receive Wi-Fi router RSS signals, which are then transmitted to a computer’s serial port. The raw data is obtained on a computer via terminal software and stored in CSV files.
The data was collected using the ESP32 mounted on the tripod in the two environments. Only data coming from the router within the scenario were stored. The dataset was captured in four directions (Up, Down, Right, and Left) in each location in two distinct sets: Reference Points (RPs) and Testing Points (TPs).
In the dataset collection procedure, each RP’ data values were recorded for 5 min in each direction, i.e., a total of 20 (5×4) min of data per point. The TP’ were recorded for 2 min each, i.e., a total of 8 (2×4) min per point.
The two selected scenarios were complex and the number of points to survey was high. Therefore, data were recorded on alternate days, with some furniture items being shifted from one room to another. Two individuals’ movements were recorded in reference and test data points. While the reference points were arbitrarily assigned symmetrically to capture the most critical information with the fewest possible efort in both scenarios, most of the test points were distributed in a grid distribution of ≈0.6 m and ≈0.75 m for the Urban Flat and Laboratory, respectively.
We have collected two datasets, one considering an urban flat FLAT and one considering a laboratory setting URBAN. Each dataset consists of two csv files: ↬DATASET_SingleRSSI_TRAIN.csv: This file contains the reference data (or radio map) collected with the ESP32 to train the positioning models. Each row represents a new sample (fingerprint) where the position is provided in columns 1 and 2 and the third column contains the RSS value. The file structure is described in Fig. 1 with an excerpt of data included in the ifle FLAT_SingleRSSI_TRAIN.csv. ↬DATASET_SingleRSSI_TEST.csv: This file contains the evaluation data collected statically in several evaluation points with the ESP32 to test the positioning models in a similar fashion than reference/training data was collected. Each row represents a new sample (fingerprint) where the position is provided in columns 1 and 2 and the third column contains the RSS value.
The main features of the two datasets are included in Table 2, where the number of samples for training and testing (‖ ‖ and ‖ℰ ‖), the number of RPs (‖ℛ ‖ ), the number of TPs (‖ ‖ ), as well as the average number of samples (measurements or S-AP fingerprints) per point and percentage of missing data/measurements ( and ℰ), are provided. (mean as point colour, the standard deviation of the valid RSS measurements and the percentage of missing data) are provided for each the RP in the two environments. The extrapolation used is a Gaussian Process Regression with a squared exponential kernel function. This figure is included to show dificulty in positioning just with a single
RSS measurement.
3.1. The -NN algorithm Wi-Fi fingerprinting based on the -Nearest Neighbors ( -NN) algorithm was introduced by Bahl and Padmanabhan in 2000, since then it has been widely used because its simplicity and smooth integration with RSS values [26].
-NN is a generic non-parametric supervised learning method introduced in 1951 [27], which is not only devoted to indoor positioning but also to gene selection [28], text classification [ 29], image classification [ 30], EEG classification [ 31], among others. -NN may be implemented as a data classifier or regressor, where the output is a class membership or the average of the values of nearest neighbours, respectively.
For Wi-Fi fingerprinting, including S-AP, -NN provides a position estimate by selecting the closest (most similar) samples from the training dataset to the operational sample. Then, the geometric centroid (a simple average) is applied to the location of the set of closest samples.
To compute the set of closest samples, a distance metric (e.g. Euclidean Distance) or a similarity metric (e.g. Cosine Similarity) is commonly used [32, 33, 34, 35]. As a fingerprint in S-AP Wi-Fi ifngerprinting is just one value, most of these distance functions and similarity metrics are equivalent. For simplicity and compliance of our algorithms with other multi-AP datasets, we assume that City Block distance (i.e. the absolute diference between two values) is appropriate for our datasets. 3.2. Analysis of in -NN As a use case, we analyse the impact of the value of in -NN for the two datasets. As mentioned before, we assume a City Block distance to compare the two values (i.e. the absolute diference between two RSS values). The mean positioning error (i.e., the Euclidean distance between the current and estimated positions) for ∈ [1, … , 100] in both datasets are provided in Fig. 3. 0 10 20 30 40
50 k value for k-NN 60 70 80 90 100
According to Fig. 3, the mean positioning error is significantly reduced over the first iterations. 8 7 )6 m ( rrro5 E ig4 n n io3 it s o P2 1 0 mean 25th perc. 50th perc. 75th perc. 90th perc. 95th perc. 99th perc.
Thus, Table 3 provides advanced error statistics (see [36, 37, 38]) on both datasets with = 100 , which can serve as a baseline for new developments. In addition, we provide the performance of the individual positioning errors as a CDF plot for = 1 and = 100 in both datasets (see Fig 4).
Empirical CDF
FLAT k = 1 FLAT k = 100 LAB k = 1 LAB k = 100 1 0.9 0.8 0.7 0.6 ) (x0.5 F 0.4 0.3 0.2 0.1 0 0 2 4
6 8 Positioning error (m) 10 12 14
Given the challenge of positioning with a single measurement from a single AP, the results obtained with mean errors of 2.58 m and 4.62 m introduce a challenge for further research.
Single-AP (S-AP) is gaining popularity given that most urban homes are not extensive and are covered by only one AP that has partial LOS in several places and, therefore, can be trusted.
Most of the existing Wi-Fi datasets either cover large environments (with hundreds of APs) or they do not resemble urban/residential medium-size flats.
This paper introduces two new datasets collected in two diferent places (an urban flat and a medium-sized laboratory) to fill this gap in the literature. Despite the dimensions of the operational areas in both scenarios and the information is limited to just one single AP, the simple baseline method based on -NN provides an accuracy between 2.85 m to 4.62 m with = 100 .
R. Kumar and V. Kumar Chaurasiya acknowledges funding from the Ministry of Education, GoI, that has supported this work. Data 6 (2021). URL: https://www.mdpi.com/2306-5729/6/6/62. doi:10.3390/data6060062. [25] N. Saccomanno, A. Brunello, A. Montanari, What you sense is not where you are: on the relationships between fingerprints and spatial knowledge in indoor positioning, IEEE Sensors Journal 22 (2022) 4951–4961. doi:10.1109/JSEN.2021.3070098. [26] P. Bahl, V. Padmanabhan, RADAR: an in-building RF-based user location and tracking system, in: Proceedings IEEE INFOCOM 2000, volume 2, 2000, pp. 775–784 vol.2. doi:10. 1109/INFCOM.2000.832252. [27] E. Fix, J. L. Hodges, Discriminatory analysis. nonparametric discrimination: Consistency properties, International Statistical Review 57 (1989) 238–247. URL: http://www.jstor.org/ stable/1403797. [28] C.-P. Lee, W.-S. Lin, Y.-M. Chen, B.-J. Kuo, Gene selection and sample classification on microarray data based on adaptive genetic algorithm/k-nearest neighbor method, Expert Systems with Applications 38 (2011) 4661–4667. URL: https://www.sciencedirect.com/science/ article/pii/S0957417410006810. doi:https://doi.org/10.1016/j.eswa.2010.07.053. [29] S. Tan, An efective refinement strategy for knn text classifier, Expert Systems with Applications 30 (2006) 290–298. URL: https://www.sciencedirect.com/science/article/pii/ S0957417405001570. doi:https://doi.org/10.1016/j.eswa.2005.07.019. [30] K. Huang, S. Li, X. Kang, L. Fang, Spectral–spatial hyperspectral image classification based on knn, Sensing and Imaging 17 (2015) 1. URL: https://doi.org/10.1007/s11220-015-0126-z. doi:10.1007/s11220-015-0126-z. [31] M. N. A. H. Sha’abani, N. Fuad, N. Jamal, M. F. Ismail, kNN and SVM Classification for
EEG: A Review, in: InECCE2019, Springer, Singapore, 2020, pp. 555–565. [32] G. Caso, L. De Nardis, M.-G. Di Benedetto, A mixed approach to similarity metric selection in afinity propagation-based wifi fingerprinting indoor positioning, Sensors 15 (2015) 27692–27720. [33] J. Torres-Sospedra, R. Montoliu, S. Trilles, Ó. Belmonte, J. Huerta, Comprehensive analysis of distance and similarity measures for wi-fi fingerprinting indoor positioning systems, Expert Systems with Applications 42 (2015) 9263–9278. [34] G. Minaev, A. Visa, R. Piché, Comprehensive survey of similarity measures for ranked based location fingerprinting algorithm, in: Int. Conf. on Indoor Positioning and Indoor Navigation (IPIN), IEEE, 2017, pp. 1–4. [35] A. Brunello, A. Montanari, N. Saccomanno, A genetic programming approach to wifi ifngerprint meta-distance learning, Pervasive and Mobile Computing 85 (2022) 101681. [36] ISO, Real time locating systems – Test and evaluation of localization and tracking systems (ISO1835), 2016. URL: https://www.iso.org/standard/62090.html. [37] F. Potortì, A. Crivello, P. Barsocchi, F. Palumbo, Evaluation of indoor localisation systems: Comments on the ISO/IEC 18305 standard, in: Int. Conf. on Indoor Positioning and Indoor Navigation (IPIN), IEEE, 2018, pp. 1–7. [38] F. Potortì, J. Torres-Sospedra, D. Quezada-Gaibor, A. R. Jiménez et al., Of-line evaluation of indoor positioning systems in diferent scenarios: The experiences from ipin 2020 competition, IEEE Sensors Journal 22 (2022) 5011–5054. doi:10.1109/JSEN.2021.3083149. [39] R. Kumar, J. Torres-Sospedra, V. Kumar Chaurasiya, Datasets for Indoor Positioning with Single-AP Wi-Fi Fingerprinting, Zenodo (2023). URL: https://doi.org/10.5281/zenodo. 7949032.
AP Access Point GNSS Global Navigation Satellite System -NN -Nearest Neighbors LBS Location Based Services LOS Line-of-Sight RNSS Regional Navigation Satellite System RP Reference Point RSS Received Signal Strengh S-AP Single-AP TP Testing Point
The dataset and sources for generating the reports and figures are available in: • Zenodo Package [39]