Line-Augmented cuVSLAM: Robust Visual Odometry in Low-Texture Environments

Can a camera-based localisation system for mobile robots stay accurate in featureless warehouse aisles by tracking straight edges as well as corner points, without slowing or replacing the existing estimator?

Visual odometry tracks small textured points from frame to frame. Warehouses and factories are full of plain walls, repetitive racking and shiny floors, so point trackers find few good matches and the position estimate drifts. That drift decides how far an autonomous mobile robot can travel before it must relocalise, which is one reason many fleets still pay for lidar.

Point-line systems such as PL-SLAM, PL-VIO and PL-VINS already add line features, but their line landmarks are deleted once the frame that created them leaves the optimisation window, a new line is never fused across several views, and points and lines share one fixed weighting regardless of how many points survive. cuVSLAM itself, NVIDIA's CUDA-accelerated SLAM that runs on Jetson boards and supports up to 32 cameras, tracks points only.

The authors add a line path that runs beside cuVSLAM's unchanged point path. Lines are detected with LSD, matched with the Line Band Descriptor and stored in a persistent registry, so an edge seen early can be reused much later. Each line's 3D estimate is a trimmed mean of its recent reconstructions, discarding the worst-agreeing ones. Line errors enter the same per-frame pose solve and the same sliding-window bundle adjustment as points, and a clamped weight gives lines more say as point observations become scarce. Switched off, the system is the original point pipeline.

In RGB-D mode with strict alignment and no scale correction, mean absolute trajectory error on all 22 TUM RGB-D freiburg3 sequences fell from 0.1960 m to 0.1205 m (38.5%) and the median from 0.1436 m to 0.0511 m (64.4%); 19 of 22 sequences improved, with the largest gains on large_cabinet (79.2%) and walking_static (88.6%). On the ten-sequence subset used in cuVSLAM's own paper the error fell from 0.1112 m to 0.0357 m (67.9%), below the published ORB-SLAM3 figure of 0.067 m (46.7% lower, MEKAWIRE calculation). On the 8 Replica room and office sequences mean translation error fell 30.1% and median rotation error 66.4% (0.580 to 0.195 degrees). Translation improved on 26 of 30 sequences and no run lost tracking.

This is a preprint evaluated on two public datasets only, one recorded with a handheld RGB-D camera and one rendered without sensor noise; there is no test on a real warehouse robot, no stereo or inertial configuration and no reported runtime or Jetson load for the added line path. Frame-to-frame precision gets slightly worse (mean one-frame translation error 0.0108 m to 0.0136 m on TUM, about 26% higher, MEKAWIRE), and errors grow on close-range views of untextured surfaces, where the point-count rule hands too much weight to unreliable repetitive edges (0.3889 m to 0.7386 m on one sequence). The authors name per-observation uncertainty and direct line tracking as the fixes.