Source-linked AI summary
Recurrent Neural Networks For Accurate RSSI Indoor Localization
Minh Tu Hoang, Brosnan Yuen, Xiaodai Dong, Tao Lu, Robert Westendorp, Kishore Reddy
TL;DR
WiFi RSSI fingerprint localization must handle unstable measurements, device differences, and spatially ambiguous fingerprints. The paper uses recurrent neural networks to predict trajectories from sequential RSSI, with filtering for temporal fluctuations; experiments report 0.75 m average error and 80% of errors under 1 m.
Problem
WiFi RSSI fingerprinting is practical but affected by instability, device heterogeneity, spatial ambiguity, and limited RSSI collection time per location.
Method
The paper develops RNN architectures that use sequential RSSI measurements and trajectory information, with filtered inputs and sequential location outputs.
Results
0.75 m average localization error with 80% of errors under 1 m was achieved by the LSTM structure.
Takeaways & Limitations
The proposed RNN approach outperformed feedforward neural networks and conventional KNN, Kalman filter, and probabilistic methods in the reported experiments.
Abstract
from arXiv · showhide
This paper proposes recurrent neuron networks (RNNs) for a fingerprinting indoor localization using WiFi. Instead of locating user's position one at a time as in the cases of conventional algorithms, our RNN solution aims at trajectory positioning and takes into account the relation among the received signal strength indicator (RSSI) measurements in a trajectory. Furthermore, a weighted average filter is proposed for both input RSSI data and sequential output locations to enhance the accuracy among the temporal fluctuations of RSSI. The results using different types of RNN including vanilla RNN, long short-term memory (LSTM), gated recurrent unit (GRU) and bidirectional LSTM (BiLSTM) are presented. On-site experiments demonstrate that the proposed structure achieves an average localization error of $0.75$ m with $80\%$ of the errors under $1$ m, which outperforms the conventional KNN algorithms and probabilistic algorithms by approximately $30\%$ under the same test environment.
I. INTRODUCTION
WiFi RSSI fingerprinting is practical but challenged by signal instability, device heterogeneity, spatial ambiguity, and limited measurements per location. The paper proposes comprehensive RNN solutions that exploit sequential RSSI and trajectory information and compares them with neural and conventional localization methods.
- RSSI is widely used because most WiFi receivers can obtain it easily.
- RSSI fingerprints are unstable because of fading and multipath effects, while different devices can report different RSSIs at the same position.
- Probabilistic methods depend on RSSI distribution assumptions or large datasets, whereas deterministic methods compare measurements with stored fingerprints.
- Existing approaches face spatial ambiguity, RSSI instability, and short RSSI collection times per location.
- The paper studies vanilla RNN, LSTM, GRU, BiRNN, BiLSTM, and BiGRU across in-house and UJIIndoorLoc datasets against neural and conventional baselines.
- Prior neural localization methods report results ranging from 82.47% classification accuracy to a 2.82±0.11 m localization accuracy, with experiments differing in APs, RPs, testing points, and grid size.
III. RNN METHODS
The proposed localization system uses recurrent networks to process sequential RSSI measurements together with trajectory information. It collects fingerprints offline and predicts locations online from mobile RSSI sequences.
- A. Recurrent Neural Network Overview: RNN outputs depend on both current inputs and historical data, making them suitable for sequential RSSI and continuous user trajectories.
- A. Recurrent Neural Network Overview: Vanilla RNN, LSTM, GRU, BiRNN, BiLSTM, and BiGRU provide alternative recurrent architectures for modeling localization sequences.
- B. Proposed Localization System: During offline training, RSSI readings collected at predefined reference points are stored as fingerprint vectors associated with physical locations.
- B. Proposed Localization System: During online testing, fewer RSSI scans collected while the user moves are processed by the proposed RNN localization architecture.
1) Data Filter:
The data-filtering and trajectory-generation pipeline addresses fluctuating RSSI measurements before recurrent localization. It smooths measurements and constructs physically plausible training trajectories from reference-point locations.
- Data Filter: RSSI measurements fluctuate because of human blocking, device movements, interference, and receiver antenna orientation.
- Data Filter: At one test position, 100 consecutive readings had a 5.5 dB standard deviation and 5% were undetected.
- Data Filter: Filtered RSSIs are used to generate random training trajectories whose consecutive locations obey a practical maximum travel distance.
- Data Filter: Trajectory generation computes Euclidean distances between locations and forms a probabilistic map for selecting the next trajectory location.
- Data Filter: The described model variants include MISO with multiple RSSI inputs and a single location output, alongside multi-time-step processing.
Model 5: P-MIMO
Model 5, P-MIMO, uses multiple RSSI readings and previously predicted locations to produce multiple output locations across time steps.
- P-MIMO takes multiple RSSIs and multiple previous predicted locations as inputs, producing multiple output locations.
- The model represents a multi-step trajectory with inputs and outputs spanning time steps 1 through T.
- Its trajectory-generation component calculates location probabilities relative to the most recent location using a Gaussian distribution.The previous location is the Gaussian mean, while σ is the standard deviation; equal-distance locations receive equal probabilities.
3) Proposed RNN Models:
The proposed RNN models learn from consecutive trajectory locations and RSSI measurements, with architectures differing in input history and output structure.
- The RNN is trained on consecutive trajectory locations so hidden states can exploit temporal correlation across RSSI measurements.Trajectory length defines memory length T, which affects both retained historical information and accumulated localization errors.
- MISO uses multiple previous RSSI readings to predict one location at time step T, while A-MISO also receives previous ground-truth locations.
- MIMO, A-MIMO, and P-MIMO map multiple time-step inputs to multiple output locations.
- A-MIMO uses ground-truth locations during training, whereas P-MIMO uses locations predicted at previous time steps.During testing, A-MISO and A-MIMO also use predicted previous locations as inputs.
- Training minimizes Euclidean distance between predicted output locations and target locations using backpropagation.Single-output and multiple-output architectures use corresponding loss formulations.
4) Sliding Window Averaging:
Sliding-window averaging combines repeated predictions of the same target location to reduce variation caused by different amounts of historical information.
- Multiple-output models produce the target location at several time steps, each using a different amount of preceding history.
- The sliding window averages repeated predictions of the target location to increase localization accuracy.
- The final target-location estimate is the average of the output set collected across previous time steps.
IV. DATABASE AND EXPERIMENTS
Experiments evaluate the proposed models in a WiFi environment using robot-collected RSSI trajectories, filter comparisons, and trajectory-based analyses.
- Experiments took place in a 21 m by 16 m area containing three corridors and 11 RSSI readings from six access points per scan.
- A mobile-device-equipped autonomous robot collected training and testing fingerprints while navigating predefined locations.
- Results were evaluated using 10-fold tests with 365,000 random training trajectories.
- A. Filter Comparison: The weighted filter reduced P-MIMO’s maximum error from 4.5 m to 3.25 m and placed 80% of errors within 1.25 m instead of 1.5 m.
- A. Filter Comparison: The filter improved localization accuracy by approximately 15% across A-MISO, MIMO, and A-MIMO, leading to its adoption for all proposed models.
B. Model Comparison
P-MIMO provides the strongest performance among the proposed models, while trajectory-based analysis explains how sequential fingerprints reduce location ambiguity.
- Model performance: 0.75±0.64 m is P-MIMO’s average localization error, outperforming MIMO at 0.80±0.67 m and MISO at 1.05±0.78 m.P-MIMO also limits the maximum error to 3.25 m, compared with 4 m for MISO.
- Memory length: 10 time steps provides slightly better accuracy, with a maximum error of 2.9 m versus 3.5 m for 40 steps and 3.75 m for 5 steps.The three memory-length settings have comparable overall performance.
- Ambiguity definition: Locations are classified as ambiguous when physically non-neighboring fingerprints have Pearson correlation above the approximately 0.9 threshold.Physical neighbors are locations separated by no more than the grid size.
- Memory length: Training trajectories longer than 8 locations eliminate ambiguity in the database, supporting a memory length of 10 locations.Single-point prediction produces an average of 27 ambiguous locations.
2) Number of Hidden Layers and Neurons:
The P-MIMO study examines architecture size, training data, optimization, and dropout to identify settings associated with lower localization error.
- Number of hidden layers and neurons: 0.75±0.64 m is the best error for a two-hidden-layer model with 100 neurons per layer, compared with 0.85±0.72 m for the best one-layer model.Increasing neuron count does not consistently improve the one-hidden-layer model.
- Training samples: 10^4 training trajectory samples is the minimum reported amount for mitigating overfitting in P-MIMO.The learning curve relates training and validation errors to the number of trajectory samples.
- Training epochs: Approximately 1,000 training epochs are reported in the learning-curve analysis.
- Learning rate and optimization: 0.75±0.64 m is achieved with ADAM at a learning rate of 0.001, the best result among the tested optimization settings.
- Dropout rate: 0.92±0.73 m is reached at a dropout rate of 0.4, while performance is mostly unchanged at dropout rates of 0.1 and 0.2.The corresponding errors are 0.72±0.68 m at 0.1 and 0.75±0.64 m at 0.2.
D. RNNs Comparison
Among the tested recurrent architectures, LSTM performs best, and its trajectory-based input outperforms memoryless and conventional localization methods in the reported comparisons.
- RNN architecture comparison: 0.75±0.64 m is LSTM’s average error, compared with 1.05±0.77 m for vanilla RNN, 0.80±0.70 m for GRU, and higher errors for the bidirectional models.LSTM and GRU keep 80% of errors within 1.2 m, while RNN and BiLSTM require 1.5 m.
- RNN architecture comparison: Gated memory helps LSTM and GRU learn long-term dependencies at T = 10, whereas vanilla RNN faces vanishing gradients at large T.LSTM slightly exceeds GRU because GRU is a simpler gated architecture with fewer memory-control features.
- RNN architecture comparison: Bidirectional models are unfavorable here because they use future information although only the first trajectory location is assumed known.The unavailable ground truth of the last location introduces error into bidirectional predictions.
- Literature comparison: P-MIMO LSTM limits 80% of errors to 1.2 m, versus 2.7 m for MLP and 2 m for the reported conventional methods.Its maximum error is 3.4 m, compared with 5.5 m for MLP and larger maxima for RADAR, Kalman filter, and Kernel method.
- Practical challenges: Sequential measurements reduce spatial ambiguity, while RSSI instability and short collection times create diverse readings at one location.The paper presents LSTM as addressing these three WiFi indoor-localization challenges.
2) Training Time Requirement:
The proposed RNN approach improves localization under speed variation and remains robust to errors in historical positions, but requires substantial offline training and leaves calibration issues for future work.
- Training Time Requirement: Approximately 4,000 s (about 1 hour 6 minutes) are required to train the optimal RNN configuration.The reported setting uses 104 training trajectory samples and approximately 1,000 epochs, at about 4 s per epoch.
- Training Time Requirement: More access points can improve performance but also increase computational cost and extend training time.The RNN models learn environmental RSSI characteristics offline before testing, so training and testing distributions may differ.
- Impacts from speed variation: 0.85±0.75 m: LSTM error remains stable as maximum user speed increases from 0.5 m/s to 2.5 m/s.SRL-KNN degrades to above 3 m once maximum speed exceeds 1.5 m/s.
- Impacts from Different Time Slots: Stationary periods can provide anchor points for recalibration, while additional sensors may improve trajectory-location estimation; detailed study remains future work.The paper notes that stationary time can exceed 80% for most mobile users.
- Stability and Robustness: Historical-position accuracy affects P-MIMO LSTM stability because previous positions are used to estimate the current location.The robustness experiment adds Gaussian error to the historical coordinates before evaluating localization.
- Stability and Robustness: Within 2 m historical-data standard deviation, localization is mostly similar to the ideal-history case, with 80% of errors at 1.5 m.The reported maximum error is 3.5 m in this scenario.
G. Other Database Comparison
On the UJIIndoorLoc database, P-MIMO LSTM achieves lower average localization error than RADAR, MLNN, Kalman filter, and MLP across two buildings.
- Dataset comparison: The paper reports consistent LSTM effectiveness on the published UJIIndoorLoc dataset despite its different grid size.Two phone users and two buildings are used, and the authors emphasize relative accuracy comparisons.
- Average localization errors: 4.2±3.2 m: P-MIMO LSTM’s average error across both UJIIndoorLoc buildings is lower than all listed comparison methods.The corresponding averages are 8.1±4.9 m for RADAR, 7.5±3.8 m for MLNN, 8.2±4.7 m for Kalman filter, and 8.2±5.2 m for MLP.
- Error distribution: 11.5 m: P-MIMO LSTM has the smallest reported maximum localization error among the compared methods.The maximum errors are 22 m for MLNN and 28 m for RADAR.
- Error distribution: 80% of P-MIMO LSTM errors are below 7 m, compared with 12 m for MLP, MLNN, and RADAR.The comparison uses the localization-error CDF for the UJIIndoorLoc database.