Why Do Loop Closure and Relocalization in Indoor Laser SLAM Mistake Locations?

Figure 1 | The top shows the actual scanned side view, the bottom compares two representations. The matrix is a schematic of the method, not performance statistics.

1. The ceiling is stable, but may not distinguish locations

The classic Scan Context (SC) projects point clouds onto a polar grid: rings by distance, sectors by azimuth, and each cell stores only one value, namely themaximum height. This representation is simple and efficient, and can also handle orientation differences through sector circular shifts.

In indoor environments, ceilings, beams, or pipelines may extend continuously over long distances, while doorways, partitions, corners, equipment, and furniture tend to appear at lower heights. If each cell only retains the highest point, differences at lower levels may be omitted from the descriptor. Consequently, two different rooms or corridors could end up with similar descriptors.

The following figure from the paper is from the same frame of measured query point cloud. It contains 467 valid cells, with the highest points of 431 cells in SC coming from the upper structure; among them, 197 cells also have mid‑low echoes, but they are not retained in the single‑track description. This frame was selected according to the fixed rule of “the most ceiling‑dominated cells” to illustrate the issue and does not represent the proportion for the entire dataset.

Figure 2 | Top-down view of the queried point cloud. The orange dot is the scan origin, the original image retains a 10 m scale.

Figure 3 | The three schematic diagrams have disks of equal size, without enlarging the single‑track diagram or compressing the dual‑track diagram. Red dots indicate low‑lying cells obscured by the highest value of the single‑track; both tracks continue to use the original paper’s unified height color scale.

UpDown‑SC still retains the ceiling‑level structures, while providing separate passages for the mid‑low geometry. The ceiling offers a stable outline, and the low‑level structures help differentiate adjacent areas.

This short video converts the same frame scan from a side view to a top‑down description, making it easy to quickly see how the upper and lower paths appear.

Video 1 | From measured point cloud to upper and lower dual‑path description (21 seconds)

2. How to Keep Two Descriptions Comparable

When comparing descriptors across time and installation states, a consistent vertical reference is required. UpDown‑SC aligns the scan with the gravity direction, eliminating roll and pitch variations; different installation heights are expressed in a unified coordinate system by the measured origin height. Yaw differences are handled later by sector alignment.

The upper and lower boundaries are estimated from the height distribution of the prior map and then reused in the query scan. Below the boundary surface, the highest points of low or middle structures are retained; above the boundary surface, the lowest points of high structures are retained, respectively recording the upper edge of the lower structure and the lower edge of the upper structure. This is the “dual envelope” representation.

Each channel has its own validity mask. Matching only compares cells that are valid in both the map and the query, distinguishing between “no observation” and “observed at low altitude.” Both paths jointly participate in candidate ranking, and sector cyclic shifts are used to estimate yaw differences.

3. Retrieve and propose candidates, ICP verifies geometric relationships

After gravity alignment and dual‑path encoding of the query scan, retrieve candidate map keyframes and yaw estimates. ICP then uses the candidates as initial values to optimize the six‑degree‑of‑freedom pose of the point cloud relative to the prior map and checks geometric fitting. Verified results can be used for loop closure or global relocalization.

Figure 4 | Paper workflow illustration, read in order. The fourth panel shows the point cloud overlaid after retrieval-oriented initialization, not the result of a converged ICP; the actual registration playback can be seen in the video below.

This requires distinguishing two metrics:Recall@1 checks whether the top retrieval candidate corresponds to the correct location, while the ICP fitting acceptance rate checks if the registration meets the predefined verification criteria. The final pose accuracy still needs to be compared with the reference ground truth.

The following snippet shows continuous registration on the prior map. Gray is the fixed map, cyan is the query point cloud after ICP registration, and the red crosses mark attempts that failed verification. In the full demonstration, 91.8% corresponds to the ICP fitting acceptance rate for the candidate windows, and should be read together with the retrieval results below.

Video 2 | Continuous point cloud registration on the prior map (22.5 seconds)

4. Experimental Results

The public project page uses a unified gravity‑aligned frontend, denoted as +G. Below are excerpts of six sets of retrieval results for SC+G and UpDown‑SC+G, all counted before ICP, expressed as percentages. R@1 checks whether the top candidate is correct, R@5 checks whether the correct location is among the top five candidates; Bold the higher value in each row, and bold the average as well.

Top-1 Recall Rate R@1 (%)