OTPL-VIO: Robust Visual-Inertial Odometry with Optimal Transport Line Association and Adaptive Uncertainty

This paper presents OTPL-VIO, a robust stereo visual-inertial odometry system that enhances performance in low-texture and illumination-challenging environments by employing a training-free deep descriptor with entropy-regularized optimal transport for line association and introducing adaptive uncertainty weighting to stabilize estimation.

The Gist

Imagine you are trying to navigate a dark, empty warehouse while wearing a blindfold, but you have a friend (the camera) and a sense of balance (the IMU) to help you. This is essentially what a robot does when it tries to figure out where it is using Visual-Inertial Odometry (VIO).

Most robots rely on "point features"—like distinct corners of a box or a unique pattern on a wall—to know where they are. But what happens if you walk into a long, white hallway with no decorations (low texture) or if the lights suddenly flicker on and off (abrupt illumination changes)? The robot's "eyes" go blind because there are no unique corners to grab onto. It starts to guess, and those guesses get worse and worse, causing the robot to get lost.

This paper introduces OTPL-VIO, a new navigation system designed specifically to solve this "blind in a white room" problem. Here is how it works, explained through simple analogies:

1. The Problem: Relying Only on "Dots"

Traditional systems are like a person trying to navigate a room by only looking at dots on the floor. If the floor is clean and has no dots, or if the lights go out, the person is stuck.

• The Flaw: Even if you try to use lines (like the edges of a door frame) to help, most current systems try to find lines by first finding dots along those lines. If the dots disappear because of bad lighting, the line connection breaks, and the system fails.

2. The Solution: Giving Lines Their Own "ID Card"

The authors realized that lines (like the edge of a table or a wall corner) are everywhere, even in boring, white rooms. But to use them, the robot needs to recognize them without relying on dots.

• The "Deep Descriptor" (The ID Card):
  Imagine every line segment gets its own unique ID card or fingerprint. Instead of looking for dots on the line to identify it, the system scans the entire line and creates a summary of its "vibe" (texture, shape, context).
  • The Magic: This ID card is created automatically by a smart AI that doesn't need extra training. It's like a security guard who can instantly recognize a person's face even if they are wearing a hat or standing in the dark, just by looking at their overall silhouette.
  • Adaptability: If a line is near lots of dots, the ID card focuses on the dots. If the dots are gone, the ID card focuses entirely on the line's shape. It's a chameleon that changes its strategy based on what's available.

3. The Matching: The "Global Seating Plan" vs. "Guessing Neighbors"

Once the robot sees lines in the current frame and the previous frame, it needs to match them up.

• Old Way (Local Matching): Imagine trying to find your friend in a crowd by only looking at the person standing immediately next to you. If your friend moves or the crowd shifts, you might grab the wrong person. This is what older systems do; they get confused easily.
• New Way (Optimal Transport): The new system acts like a smart event planner. Instead of just looking at neighbors, it looks at the entire room. It asks, "Given all the lines I saw before and all the lines I see now, what is the single best way to pair them up so everyone is happy?"
  • It uses a mathematical concept called Optimal Transport (think of it as the most efficient way to move furniture from one room to another).
  • It can handle "ghosts" (lines that disappeared) and "newcomers" (lines that appeared) without panicking. It ensures that even if the view is blurry or partial, the connections remain consistent.

4. The Safety Net: "Trust but Verify"

Not all lines are created equal. A short, fuzzy line is less reliable than a long, sharp one.

• Adaptive Weighting: Imagine you are driving a car. If you see a road sign that is clear and far away, you trust it 100%. If you see a sign that is blurry and close, you trust it less.
• The system does the same thing. It calculates how "noisy" or unreliable a line is. If a line is shaky or short, the system turns down the volume on that clue so it doesn't mess up the robot's position. If the line is strong, it turns up the volume. This prevents bad data from dragging the robot off course.

The Result: A Robot That Never Gets Lost (Even in the Dark)

The authors tested this system in:

1. Standard labs: Where it beat all other robots in accuracy.
2. Harsh environments: Where lights flickered wildly and walls were blank.
3. Real-world scenarios: Like walking through a dimly lit warehouse with sudden bright lights.

The Verdict:
While other robots stumbled, tripped, or got lost in these confusing environments, OTPL-VIO kept walking straight. It did this by:

1. Giving lines their own unique "fingerprint" so they don't need dots to be recognized.
2. Using a "global seating plan" to match lines perfectly, even when the view is messy.
3. Ignoring bad clues and listening only to the reliable ones.

It's like upgrading from a robot that needs a map with every single streetlight to a robot that can navigate by the shape of the buildings themselves, even if the streetlights go out.

Technical Summary

Here is a detailed technical summary of the paper OTPL-VIO: Robust Visual-Inertial Odometry with Optimal Transport Line Association and Adaptive Uncertainty.

1. Problem Statement

Robust Visual-Inertial Odometry (VIO) faces significant challenges in environments characterized by low texture and abrupt illumination changes.

• Limitations of Point Features: In low-texture scenes, repeatable keypoints are scarce. Under rapid lighting changes, photometric-based matching becomes unstable, leading to sparse features, ambiguous associations, and triangulation failures.
• Limitations of Existing Line Systems: While line structures offer complementary geometric constraints, many existing point-line systems rely on point-guided line association (using keypoints along a line to match the line). This approach fails when point support is weak or unstable. Furthermore, traditional handcrafted line descriptors (e.g., LBD) lack discriminability in repetitive or low-gradient regions, and heavy learning-based methods often incur computational costs that hinder real-time deployment.
• Optimization Instability: Line measurements often exhibit heterogeneous reliability (e.g., short segments are noisier than long ones), yet many systems apply uniform weighting, leading to biased constraints and unstable pose estimation.

2. Methodology: OTPL-VIO

The authors propose OTPL-VIO, a stereo point-line VIO system that integrates a lightweight learning-based front-end with a factor-graph back-end. The system consists of three core innovations:

A. Lightweight Deep Line Descriptor (Training-Free)

Instead of training a new network, the system leverages features from an existing unified point-line network (PL-Net).

• Feature Aggregation: For each detected line segment, the system samples $N_s$ points along the segment in the feature map coordinates.
• Dual-Branch Pooling: It extracts features from both the line-specific branch ($F_{line}$) and the point-specific branch ($F_{pt}$) of the network.
• Adaptive Weighting: The final descriptor is a weighted concatenation of these two branches. The weights ($\gamma_{line}$ and $\gamma_{pt}$) are determined dynamically based on the local keypoint density near the segment:
  • In point-rich areas, the descriptor relies more on the point-specific branch (leveraging local structures).
  • In low-texture areas (low keypoint density), it shifts reliance to the line-specific branch to capture holistic structural context.
• Benefit: This creates a robust, training-free descriptor that adapts to the scene's texture quality.

B. Global Optimal Transport (OT) Line Association

To solve the ambiguity in line matching under poor visual conditions, the system formulates line association as an entropy-regularized Optimal Transport (OT) problem.

• Global Consistency: Unlike local Nearest Neighbor (NN) matching, OT considers all line segments in two frames simultaneously to find the globally optimal set of correspondences.
• Handling Unmatched Segments: The formulation introduces virtual nodes on both sides of the transport matrix. This allows segments that cannot be matched (due to occlusion or ambiguity) to be "transported" to a virtual node rather than forcing a false match, effectively handling partial observations and outliers.
• Mass Definition: Line length is used as the transported mass to ensure longer, more reliable segments have higher influence.

C. Reliability-Adaptive Optimization

The system introduces a weighting mechanism to regulate the influence of line constraints during the back-end optimization (Bundle Adjustment).

• Geometric Weighting ($w_{geo}$): Based on the observation that short lines have higher directional uncertainty under endpoint noise, the weight is inversely proportional to the estimated orientation uncertainty ($\sigma^2_\theta \propto 1/L^2$). Short segments are down-weighted.
• Visibility Weighting ($w_{vis}$): Segments tracked for a sufficient number of frames are given higher weight; those with short track lengths are down-weighted.
• Result: This prevents noisy or unstable line measurements from corrupting the pose estimation.

3. Key Contributions

1. Training-Free Deep Descriptor: A novel segment-level descriptor that aggregates contextual features from a pre-trained network without additional training, utilizing adaptive weighting based on local point density.
2. OT-Based Global Association: A global matching strategy using entropy-regularized Optimal Transport that ensures consistent correspondences under ambiguity and supports unmatched segments via virtual nodes.
3. Reliability-Adaptive Weighting: A strategy that dynamically adjusts the weight of line constraints based on segment length (geometric stability) and track persistence, stabilizing the optimization process.
4. Comprehensive Evaluation: Demonstrated superior performance on standard benchmarks (EuRoC, UMA-VI) and real-world deployments involving severe illumination changes and low-texture corridors.

4. Experimental Results

The system was evaluated against state-of-the-art baselines (VINS-Fusion, ORB-SLAM3, AirSLAM, PL-SLAM, etc.) on multiple datasets:

• EuRoC Dataset: OTPL-VIO achieved the best overall Average RMSE (8.06 cm), outperforming the strongest baseline (AirSLAM, 11.18 cm) by 27.9%. The improvement was most significant in "Hard" sequences with low texture and illumination changes.
• UMA-VI (Illumination Changes): On sequences with abrupt lighting shifts, the system achieved an average RMSE of 25.5 cm, reducing error by 42.2% compared to AirSLAM. Traditional methods (like ORB-SLAM3) frequently failed or suffered large drifts.
• UMA-VI (Low Texture): In low-texture corridors, the system achieved 11.60 cm RMSE, significantly outperforming point-guided baselines (e.g., AirSLAM at 26.04 cm).
• Real-World Deployment: Tested on custom indoor routes with rapid light transitions. OTPL-VIO maintained stable trajectories where other systems experienced large drifts or tracking failures.
• Real-Time Performance: The system runs in real-time with an average latency of 32.89 ms per frame on a standard laptop GPU, faster than VINS-Fusion (42.35 ms) and AirSLAM (38.36 ms).

5. Significance

OTPL-VIO addresses a critical gap in visual odometry: the inability of current systems to maintain robustness when both point features are sparse and illumination is unstable.

• Decoupling from Point Reliability: By introducing a dedicated, adaptive line descriptor and global matching, the system no longer relies on the stability of keypoints to associate lines.
• Robustness to Degradation: The combination of OT matching and adaptive weighting allows the system to gracefully handle outliers and noisy measurements, which is crucial for autonomous navigation in dynamic, unstructured indoor environments.
• Efficiency: It proves that high robustness does not require heavy, redundant neural networks; a lightweight, training-free approach integrated with mathematical optimization (OT) can achieve state-of-the-art results in real-time.

In conclusion, OTPL-VIO represents a significant step forward in making VIO systems viable for challenging, real-world industrial and indoor applications where lighting and texture conditions are unpredictable.