Fusion of Wi-Fi and LiDAR for Large-Scale Indoor Robot SLAM
Introduction
Simultaneous Localization and Mapping (SLAM) is a fundamental technology enabling autonomous navigation for mobile robots in unknown environments. While laser-based SLAM systems offer high precision, they face significant challenges in large indoor environments where incorrect loop closures can lead to substantial pose estimation errors. This paper presents an innovative graph-based SLAM algorithm that integrates Wi-Fi fingerprint sequences with LiDAR submaps to enhance localization accuracy in expansive indoor settings.
The proposed method addresses the limitations of standalone LiDAR SLAM by leveraging the complementary strengths of Wi-Fi signals and laser scans. Wi-Fi infrastructure, widely available in indoor spaces, provides coarse but reliable area recognition, while LiDAR delivers precise geometric measurements. By fusing these two modalities, the system achieves robust loop closure detection and significantly reduces cumulative errors that degrade traditional SLAM performance in feature-poor or repetitive environments.
Methodology Overview
The algorithm operates through a systematic pipeline that begins with multi-sensor data acquisition. A mobile robot equipped with wheel encoders, LiDAR, and Wi-Fi scanners collects odometry, laser range measurements, and wireless signal strength data. These inputs feed into parallel processing streams that construct Wi-Fi fingerprint sequences and LiDAR submaps using sliding time windows.
Key innovations include a novel submap selection mechanism based on Wi-Fi fingerprint similarity analysis. Rather than exhaustively matching all possible laser submaps—a computationally expensive process prone to false positives—the system first identifies candidate loop closure regions through statistical analysis of wireless signal patterns. This hierarchical approach dramatically improves computational efficiency while maintaining high detection accuracy.
Following submap selection, feature-based matching refines potential loop closures before feeding constraints into a graph optimization framework. The final stage merges odometric constraints with verified loop closures to produce a globally consistent map and optimized trajectory.
Wi-Fi Fingerprint Sequence Processing
The system treats Wi-Fi signals as spatial fingerprints by recording MAC addresses and signal strengths from detectable access points. Each fingerprint represents a snapshot of the radio environment at a specific location. To improve reliability against signal fluctuations, the algorithm extends single fingerprints into sequences using a sliding window approach.
Sequence similarity evaluation employs cosine similarity metrics to compare pairs of fingerprint sequences. This technique measures angular similarity between signal strength vectors, effectively capturing environmental resemblances despite absolute signal variations. The system computes both mean similarity and standard deviation across all fingerprint pairs within sequences, creating a robust statistical profile for each location.
Threshold-based filtering then identifies high-probability loop closure candidates. Sequences exhibiting both high mean similarity (indicating spatial proximity) and low standard deviation (suggesting stable signal patterns) trigger subsequent LiDAR submap processing. This dual-threshold strategy effectively filters out transient signal matches while preserving genuine revisits to previously mapped areas.
LiDAR Submap Construction and Selection
Concurrent with Wi-Fi processing, the system builds local LiDAR submaps using occupancy grid mapping techniques. Each submap covers a limited spatial extent corresponding to the sliding window duration, balancing detail with computational tractability. The correlative scan matching method aligns successive laser scans within each submap, maintaining local consistency.
The Wi-Fi fingerprint analysis directly informs submap selection for loop closure verification. Only submaps associated with statistically significant Wi-Fi sequence matches undergo feature extraction and matching. This selective processing avoids unnecessary computations on dissimilar regions while focusing resources on probable loop closure candidates.
For chosen submaps, the algorithm applies AKAZE feature detection to identify distinctive environmental patterns. This scale-invariant feature transform proves particularly effective for indoor environments with geometric structures like walls, corners, and doorways. Feature matching between current and historical submaps then estimates relative pose transformations with associated confidence scores.
Graph Optimization Framework
The system represents SLAM as a pose graph optimization problem where nodes correspond to robot poses and edges encode constraints between them. Two constraint types contribute to the graph: consecutive pose links from odometry, and non-consecutive links from verified loop closures.
The g2o optimization framework solves this graph using iterative nonlinear least squares methods. Each loop closure constraint introduces a correction factor that pulls the trajectory toward global consistency, counteracting odometric drift. The optimization process automatically reweights constraints based on their estimated uncertainties, with high-confidence loop closures exerting stronger influence on the final solution.
This formulation elegantly handles the multi-sensor nature of the system. Wi-Fi-derived information indirectly contributes through selective submap matching, while LiDAR provides precise geometric constraints. The fusion occurs naturally within the graph structure, avoiding complex sensor fusion heuristics while maintaining mathematical rigor.
Experimental Validation
Performance evaluation occurred across three large-scale indoor environments spanning up to 180×80 meters, with robot trajectories exceeding 2.7 kilometers in the most extensive test. The experiments deliberately incorporated challenging scenarios including long featureless corridors, symmetrical layouts, and dynamic interference to validate robustness.
Comparative analysis against standalone LiDAR SLAM demonstrated consistent accuracy improvements across all test cases. The proposed system achieved localization errors of 0.78m, 0.67m, and 0.89m in respective datasets—representing precision enhancements between 48.6% to 68.7% over conventional approaches. These gains came despite significantly reduced computational load, with loop closure processing times decreasing by 22.6% to 40.3%.
Parameter sensitivity studies revealed optimal operating ranges for key thresholds. A Wi-Fi similarity threshold around 0.3 combined with standard deviation thresholds between 0.2-0.4 produced the best balance between detection sensitivity and false positive rejection. Window durations of 20-30 seconds optimally captured environmental features without excessive memory requirements.
Qualitative assessment of generated maps showed excellent structural consistency with minimal distortion or alignment errors. The system successfully handled challenging loop closure scenarios where geometric similarity alone would have produced incorrect matches, demonstrating the value of Wi-Fi-assisted candidate selection.
Discussion and Implications
The experimental results confirm that wireless signal patterns can effectively guide laser-based loop closure detection without requiring pre-deployed infrastructure or manual fingerprint maps. This represents a significant advance over traditional Wi-Fi positioning systems that demand labor-intensive offline calibration.
The hierarchical processing architecture offers particular advantages in large environments where computational efficiency becomes critical. By eliminating improbable loop closure candidates early in the processing chain, the system maintains real-time performance even during extended mapping sessions. This scalability makes the approach suitable for warehouse logistics, shopping mall navigation, and other large-scale applications.
Several practical considerations emerge from this work. Wi-Fi signal stability proves more important than absolute signal strength for reliable fingerprinting. Environments with dense, overlapping access points provide richer fingerprint data than sparse deployments. The system demonstrates reasonable tolerance to temporary signal dropouts through its sequence-based matching approach.
Conclusion
This research presents a robust solution to large-scale indoor SLAM by intelligently combining ubiquitous wireless signals with precise laser measurements. The fusion architecture overcomes key limitations of standalone systems—LiDAR’s susceptibility to perceptual aliasing and Wi-Fi’s positioning imprecision—through complementary sensor exploitation.
The algorithm’s modular design permits straightforward integration with existing robotic platforms while its parameterized thresholds allow adaptation to diverse environments. Future extensions could incorporate additional sensor modalities like ultra-wideband or visual features to further enhance reliability in challenging conditions.
By significantly improving pose estimation accuracy while reducing computational overhead, this work advances the feasibility of long-duration autonomous operation in complex indoor spaces. The techniques demonstrated here have immediate applications in service robotics, industrial automation, and infrastructure inspection scenarios where reliable navigation in expansive, feature-variable environments is paramount.
For further technical details, refer to the full paper at https://doi.org/10.19734/j.issn.1001-3695.2024.06.0244
Was this helpful?
0 / 0