Modeling Perceptual Aliasing in SLAM via Discrete-Continuous Graphical Models

Pierre-Yves Lajoie, Siyi Hu, Giovanni Beltrame, Luca Carlone

This paper, published in IEEE Robotics and Automation Letters (2019), tackles one of the most persistent "Achilles' heels" of robot navigation: the problem of things looking too much alike.

TL;DR

The authors introduce a Discrete-Continuous Graphical Model (DC-GM) that explicitly models "perceptual aliasing"—when a robot gets confused by similar-looking places. By using a semidefinite relaxation (SDP) to solve the resulting optimization problem, they provide a robust SLAM framework that handles highly correlated outliers and, crucially, doesn't require a "good" initial guess to succeed.

What Problem Does This Paper Address?

Simultaneous Localization and Mapping (SLAM) is how a robot builds a map of an unknown environment while keeping track of its own location. The system relies on "loop closures"—recognizing a place it has seen before to correct its estimated trajectory.

However, in the real world, robots encounter perceptual aliasing. Imagine a long office hallway with identical doors or a warehouse with repeating shelf patterns. The robot might think it’s at "Door A" when it’s actually at "Door B." This creates "outliers"—false measurements that can catastrophically warp the map.

Standard SLAM solvers struggle with two things:

  1. Correlated Outliers: Perceptual aliasing doesn't just create one random bad measurement; it often creates a whole cluster of "mutually consistent" bad measurements (e.g., several false matches along that identical hallway).
  2. The "Initial Guess" Trap: Most robust SLAM algorithms need a decent starting estimate of the robot's path to tell inliers from outliers. If the initial odometry is poor, the algorithm often fails to recover.

Methods

The authors propose a unified framework called a Discrete-Continuous Graphical Model (DC-GM):

  • Continuous Variables: These represent the standard SLAM problem—the actual 2D or 3D poses (positions and orientations) of the robot.
  • Discrete Variables: These are binary "switches" associated with each loop closure measurement. They decide whether to "accept" a measurement as an inlier or "reject" it as an outlier.
  • Correlation Modeling: Unlike previous methods that treat each measurement as independent, this model uses a Markov Random Field (MRF) structure. It explicitly models the positive correlation between outliers, essentially saying: "If these three measurements look consistent with each other but conflict with the rest of the map, they might be a cluster of aliased outliers."

To solve this complex, non-convex mix of discrete and continuous variables, the authors derive a Semidefinite Relaxation (SDP). This transforms the "hard" problem into a "convex" one that can be solved with global optimality guarantees, meaning the algorithm finds the best possible solution without needing a starting guess.

Datasets

The researchers evaluated their approach on both simulated and real-world benchmarks:

  • Synthetic Datasets: A simulated grid world and a more complex "Manhattan World" graph. These were used to test the limits of the system by injecting up to 50% correlated outliers.
  • Real-World Datasets:
    • CSAIL: A dataset from MIT (1,045 poses).
  • FR079 & FRH: Datasets from the University of Freiburg.
  • The authors "spoiled" these real datasets by adding groups of false loop closures to simulate extreme perceptual aliasing.

Key Insights and Ideas

  • Modeling vs. Mitigating: Most SLAM research tries to mitigate the effect of outliers (e.g., using "heavy-tailed" error distributions). This paper argues we should model the physical process (perceptual aliasing) that creates them.
  • The Power of Convexity: By using SDP relaxation, the authors move away from "local" search algorithms (like Gauss-Newton) which are prone to getting stuck in "local minima" (wrong solutions). This makes the system "initialization-free."
  • Physical Meaning: The parameters in their model (like the maximum admissible residual $c$) have clear physical meanings, making the system easier to tune than abstract mathematical weights used in other robust solvers.

Key Results

  • Superior Robustness: In simulations with 50% correlated outliers, the DC-GM approach successfully rejected all outliers, while competing methods like Vertigo or DCS (Dynamic Covariance Scaling) often failed or required perfect parameter tuning.
  • Real-World Performance: On the FR079 and CSAIL datasets, DC-GM achieved the lowest average translation error. For example, on CSAIL, DC-GM maintained an error of ~0.04m, while some state-of-the-art competitors jumped to over 1.4m because they couldn't distinguish the correlated outliers.
  • No Guess Required: The system successfully optimized maps starting from "zero" information, whereas traditional methods failed if the initial odometry was significantly drifted.

Limitations and Future Work

  • Computational Speed: The main drawback is that solving SDPs is computationally expensive. The current implementation (using Matlab and generic solvers) is too slow for real-time robot operation. The authors suggest that specialized solvers (like SE-Sync) could bridge this gap.
  • Dimensionality: While the math is general, the experiments were primarily focused on 2D SLAM. Extending the rigorous testing to 3D environments is a stated next step.
  • Batch vs. Incremental: The current algorithm is a "batch" solver (it processes the whole map at once). For a robot to use this live, it would need to be converted into an "incremental" solver that updates as the robot moves.

Summary updated

Abstract

Perceptual aliasing is one of the main causes of the failure for simultaneous localization and mapping (SLAM) systems operating in the wild. Perceptual aliasing is a phenomenon where different places generate a similar visual (or, in general, perceptual) footprint. This causes spurious measurements to be fed to the SLAM estimator, which typically results in incorrect localization and mapping results. This problem is exacerbated by the fact that those outliers are highly correlated, in the sense that perceptual aliasing creates a large number of mutually consistent outliers. Another issue stems from the fact that most state-of-the-art techniques rely on a given trajectory guess (e.g., from odometry) to discern between inliers and outliers, and this makes the resulting pipeline brittle, since the accumulation of error may result in incorrect choices and recovery from failures is far from trivial. This paper provides a unified framework to model perceptual aliasing in SLAM and provides practical algorithms that can cope with outliers without relying on any initial guess. We present two main contributions. The first is a discrete-continuous graphical model (DC-GM) for SLAM: The continuous portion of the DC-GM captures the standard SLAM problem, while the discrete portion describes the selection of the outliers and models their correlation. The second contribution is a semidefinite relaxation to perform inference in the DC-GM that returns estimates with provable sub-optimality guarantees. Experimental results on standard benchmarking datasets show that the proposed technique compares favorably with state-of-the-art methods while not relying on an initial guess for optimization.
BibTeX
@article{lajoie2019modeling,
  author = {Pierre-Yves Lajoie and Siyi Hu and Giovanni Beltrame and Luca Carlone},
  title = {{Modeling Perceptual Aliasing in SLAM via Discrete-Continuous Graphical Models}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2019},
  url = {https://openreview.net/forum?id=Htu5yQPgqg},
}

Loading related content...

Figure