DKDhruv Karnik

Visual-inertial state estimation

Adaptive UKF / PF for Underwater Navigation

Adaptive nonlinear filtering that varies trust in visual measurements using reprojection error as a visual-uncertainty signal.

UKF · Particle Filter · Visual-Inertial · Underwater Robotics

Overview

Underwater visual measurements can degrade because of turbidity, low light, feature loss, backscatter, and refraction. Fixed-noise estimators can continue to trust those measurements too strongly. This project evaluates adaptive methods that reduce visual trust when reprojection error rises.

Estimator set

The final comparison includes a degraded visual baseline, fixed UKF, tuned Adaptive UKF, fixed particle filter, Adaptive PF v9, and a Hybrid Adaptive UKF/PF estimator. All methods are evaluated against ground truth on the same trajectory data.

Adaptive UKFInflates visual measurement covariance as reprojection error increasesAdaptive PFBroadens the visual likelihood under higher uncertainty

Dataset and evaluation

The evaluation uses eight real underwater trajectories: fjord_1 through fjord_6 and mclab_1 through mclab_2. Position RMSE is evaluated over all samples and over high visual-uncertainty sections defined by the 80th, 90th, and 95th reprojection-error percentiles.

Results

The tuned Adaptive UKF improves over the fixed UKF across the tested trajectories. Adaptive PF v9 provides the strongest overall accuracy in most comparisons, while the Hybrid Adaptive UKF/PF stays close to the best estimator and provides an explicit switching mechanism during high visual-uncertainty intervals.

Representative all-samples RMSEfjord_1: Adaptive PF 1.34 m · fjord_3: 1.74 m · fjord_6: 2.63 mHybrid behaviorSwitching is driven by reprojection-error spikes with hysteresis to avoid rapid mode chatter

Tradeoff

The hybrid method is useful because its response to changing visual quality is interpretable, but it has the highest runtime. Adaptive PF v9 provides the strongest accuracy-to-complexity balance in the final experiments.