PL-VIWO2: Robust Visual-Inertial-Wheel Odometry

• PL-VIWO2 is a lightweight, fast, and robust visual-inertial-wheel odometry system that fuses camera, IMU, and wheel encoder data for enhanced urban vehicle localization.
• It employs a novel geometric line feature processing framework and SE(2)-to-SE(3) wheel pre-integration to efficiently handle low-texture regions and dynamic objects.
• A motion consistency check filters out dynamic features, resulting in improved localization accuracy and real-time performance in challenging outdoor urban settings.

PL-VIWO2 refers to a lightweight, fast, and robust visual-inertial-wheel odometry system designed for ground vehicle localization, specifically targeting long-term operation in complex outdoor urban environments. By fusing data from a camera (supporting monocular and stereo modes), a high-rate inertial measurement unit (IMU), and wheel encoders, PL-VIWO2 addresses degradation commonly seen in low-cost, vision-based odometry systems when encountering dynamic objects, low-texture regions, or degenerate motion patterns. Its core technical advances include a novel geometric line feature processing framework, wheel updates anchored in planar motion constraints, and an integrated motion consistency check that tightly couples all sensors for feature outlier rejection. Experimental and simulation comparisons with state-of-the-art odometry frameworks demonstrate that PL-VIWO2 yields improved localization accuracy, runtime efficiency, and robustness (Zhang et al., 25 Sep 2025).

1. System Architecture and State Representation

PL-VIWO2 operates as a tightly coupled sensor fusion system where the state vector aggregates the current IMU state—including pose, velocity, and biases—with a history of "clone" states for multi-state constraint Kalman filtering. Its three principal input modalities are:

• IMU: Provides high-rate inertial propagation, modeling the system’s pose and velocity at timescales up to hundreds of Hz.
• Wheel Encoder: Supplies planar displacement and yaw rate observations, representing the 2D motion typically exhibited by wheeled vehicles.
• Camera: Delivers key visual constraints via the extraction and tracking of both point and line features.

The system maintains the following state vector at time $k$:

$x_k = (x^I_k, x^H_k)$

where $x^I_k$ is the current IMU state and $x^H_k$ is a set of historical clones necessary for multi-frame visual and wheel constraints.

IMU integration performs high-frequency state propagation. Lower frequency updates are provided by the camera (visual features) and wheel encoder (planar displacement). This multi-modal approach allows PL-VIWO2 to resist failure cases arising from visual feature sparsity or textureless environments.

2. Line Feature Processing Framework

PL-VIWO2 introduces a streamlined and robust methodology for leveraging line features, which augments the geometric constraints provided by points, particularly in urban landscapes dominated by prevailing structural edges.

Key Components

• 2D Line Detection and Merging: The system employs a Fast Line Detector (FLD) for efficient extraction of line segments. To counteract overfragmentation (where long real-world lines are detected as numerous short segments), a merging step combines colinear, spatially contiguous fragments as in techniques similar to AirVO.
• Manhattan World Line Classification: Leveraging the prevalent "Manhattan World" assumption in urban scenes, PL-VIWO2 classifies detected lines by their orientation along the orthogonal axes (x, y, z) relative to the IMU frame. Vanishing points are computed using camera-to-IMU rotation:

$$\begin{aligned} \mathrm{vp}_x &= \pi({_I^C}R [1,0,0]^\top) \ \mathrm{vp}_y &= \pi({_I^C}R [0,1,0]^\top) \ \mathrm{vp}_z &= \pi({_I^C}R [0,0,1]^\top) \end{aligned}$$

where $\pi(\cdot)$ is the camera projection function.

• Point–Line Assignment and Two-Stage Tracking: Points are associated with nearby lines if their Euclidean distance to a segment falls below a threshold, enabling robust line tracking. The initial association enables direct tracking; for ambiguous cases, optical flow is applied to intermediate points interpolated along the segment.
• 3D Line Triangulation and Refinement: Line triangulation proceeds by (a) using two or more triangulated points located on a line, or (b) from the intersection of back-projected planes. In degenerate motion scenarios (such as pure forward motion with respect to a line), the system uses the known Manhattan direction and a single point to initialize the Plücker coordinates:

$v = {}_G^I R \cdot u, \quad n = p \times v$

with $u$ denoting the unit Manhattan direction and $p$ a 3D point on the line. A minimal 4 DoF optimization (using an orthonormal representation) refines each line's pose, minimizing reprojection and point-to-line residuals jointly.

3. SE(2)-Constrained SE(3) Wheel Pre-Integration

PL-VIWO2’s treatment of wheel encoder data is notable for formally fusing planar odometry into the 3D pose estimation framework. The angular velocities from left ($\omega_l$) and right ($\omega_r$) wheels and radii ($r_l$, $r_r$) yield 2D translational and angular velocities:

$${}^W v_k = \frac{\omega_r r_r + \omega_l r_l}{2}, \quad {}^W \omega_k = \frac{\omega_r r_r - \omega_l r_l}{b}$$

($b$ = axle length). Under a no-slip planar constraint ($v_z=0$, $\omega_x=\omega_y=0$), this motion is "lifted" from SE(2) to SE(3):

$$\bar{z}_W = \begin{bmatrix} \text{Log}(\Delta R^g_W) \ \Delta p^g_W \end{bmatrix} = \begin{bmatrix} \int_{t_{k-1}}^{t_k} {}^W\omega\, dt \ \int_{t_{k-1}}^{t_k} {}^W v \cdot ( \cos\theta, \sin\theta )\, dt \end{bmatrix}$$

This measurement is incorporated in a tightly coupled Kalman filter step, constraining 3D motion estimates—especially in degenerate scenarios—by exploiting the inherent planarity of ground vehicle motion.

4. Motion Consistency Check (MCC)

Reducing the influence of dynamic, non-static features is critical for urban odometry. The motion consistency check (MCC) module filters features by comparing each feature’s observed 2D displacement with the predicted displacement from the fused IMU and wheel motion:

$$r = \frac{1}{n} \sum_{i=1}^{n} \| z_i - \pi( {}^{C_i}_G R ( {}^G \hat{p}_f - {}^G p_{C_i} ) ) \|$$

where $z_i$ is the $i^\text{th}$ observation, $^G \hat{p}_f$ is the estimated feature position, and $^G p_{C_i}$ is the pose from IMU and wheel fusion. Features with $r$ above a threshold are identified as dynamic and excluded from the state update, improving robustness to moving objects and dynamic scene elements.

5. Empirical Performance and Comparative Evaluation

PL-VIWO2’s methods were validated both in Monte Carlo simulations and on public autonomous driving datasets such as the KAIST Urban Dataset. Key findings include:

• Triangulation Robustness: The hybrid line initialization strategy (using a single point and known direction for degenerate cases) achieves lower error than pure plane-intersection methods in simulated scenarios, particularly under challenging vehicle dynamics.
• Real-World Experiments: On the KAIST dataset, PL-VIWO2 delivers superior accuracy and robustness—especially in highway sequences where planar constraints are essential and conventional methods often degrade. In urban sequences, the combination of robust line tracking and planar wheel updates yields leading or second-best localization performance among monocular and stereo platforms.
• Runtime Efficiency: Feature extraction, classification, and matching are significantly expedited relative to earlier systems (e.g., PL-VINS), with the combined processing pipeline operating in real time on standard embedded hardware.

A summary table of comparative aspects is as follows:

Aspect	PL-VIWO2	PL-VIWO	PL-VINS / VINS-Mono
Sensor Configs	Monocular / Stereo, Wheel, IMU	Monocular, Wheel, IMU	Mono/Stereo (no wheel)
Line Processing	Lightweight, Manhattan-based	Vanishing point-based	Expensive descriptor-based
Planar Constraints	SE(2)-to-SE(3) lifting	Wheel encoder in MSCKF	No planar wheel constraints
Efficiency	Real-time, low compute	Real-time	Higher runtime (lines)

6. Practical Applications, Limitations, and Prospects

PL-VIWO2 is optimized for outdoor vehicle localization in environments suffering from low texture, strong dynamic interference, and extended planar motion. Its multimodal sensor fusion makes it especially suitable for autonomous driving and robotics where monocular vision, with its inherent scale ambiguity, is augmented by wheel odometry and inertial data.

Key limitations include:

• In highly degenerate or slippage scenarios where wheel odometry is unreliable, estimation accuracy may degrade.
• Feature association, especially for lines in highly dynamic scenes, may remain challenging and could introduce misclassifications.
• Robustness ultimately depends on tuning feature rejection thresholds in the MCC and optimizing data association in highly cluttered urban scenes.

Prospects for further enhancement mentioned in the source include advanced geometric constraints for features, improved optimization for minimal parameterizations in line modeling, and further generalization to arbitrary sensor configurations (e.g., full stereo deployments or integration of semantic information).

7. Significance and Outlook

PL-VIWO2 exemplifies current trends in robust sensor fusion for mobile robotics: exploiting lightweight geometric representations (notably, lines under Manhattan world assumptions), constraining motion models using physical insights (such as planar vehicle locomotion), and unifying disparate sensor modalities in a tightly coupled filter-based estimation framework. Its demonstrated improvements over established baselines in both accuracy and computational efficiency position it as a reference system for benchmarking in long-term, large-scale autonomous navigation scenarios (Zhang et al., 25 Sep 2025). Extension to denser 3D representations, richer feature semantics, or adaptive parameterization for highly non-urban scenes constitutes plausible avenues for future research.