Source-linked AI summary
OverlapNet: Loop Closing for LiDAR-based SLAM
Xieyuanli Chen, Thomas Läbe, Andres Milioto, Timo Röhling, Olga Vysotska, Alexandre Haag, Jens Behley, Cyrill Stachniss
TL;DR
The paper addresses LiDAR-based loop closure detection for SLAM, where loop closures are needed to correct accumulated drift. OverlapNet predicts range-image overlap and relative yaw from multiple LiDAR cues, then supports loop detection and scan matching. Experiments on KITTI and Ford campus suggest improved performance over state-of-the-art methods and generalization to unseen environments.
Problem
The paper addresses loop closing for SLAM using 3D laser scans, a task needed for global data association and correction of accumulated drift.
Method
OverlapNet is a lightweight siamese deep network using spherical LiDAR projections and multiple cues to predict scan overlap and relative yaw.
Results
Experiments suggest that combining OverlapNet with odometry outperforms state-of-the-art methods and generalizes to environments unseen during training.
Takeaways & Limitations
The approach provides loop closure candidates, improves pose estimation toward globally consistent maps, and initializes ICP with predicted yaw.
Abstract
from arXiv · showhide
Simultaneous localization and mapping (SLAM) is a fundamental capability required by most autonomous systems. In this paper, we address the problem of loop closing for SLAM based on 3D laser scans recorded by autonomous cars. Our approach utilizes a deep neural network exploiting different cues generated from LiDAR data for finding loop closures. It estimates an image overlap generalized to range images and provides a relative yaw angle estimate between pairs of scans. Based on such predictions, we tackle loop closure detection and integrate our approach into an existing SLAM system to improve its mapping results. We evaluate our approach on sequences of the KITTI odometry benchmark and the Ford campus dataset. We show that our method can effectively detect loop closures surpassing the detection performance of state-of-the-art methods. To highlight the generalization capabilities of our approach, we evaluate our model on the Ford campus dataset while using only KITTI for training. The experiments show that the learned representation is able to provide reliable loop closure candidates, also in unseen environments.
I. INTRODUCTION
The paper introduces OverlapNet, a deep network for LiDAR loop closing that predicts scan overlap and relative yaw from multiple LiDAR cues. Integrated with odometry and SLAM, it targets accurate loop candidates and generalization across environments.
- Loop closures correct accumulated drift and enable globally consistent SLAM maps.
- OverlapNet predicts range-image overlap and relative yaw between LiDAR scans using depth, normals, intensity, and semantic-class probabilities.The network uses spherical projections rather than raw point clouds.
- The experiments evaluate KITTI training and testing, ablations, online SLAM, and transfer to the differently structured Ford campus dataset.
- The approach combines overlap predictions with odometry to detect loop closures and initializes ICP for relative scan matching.
- Related work: Unlike earlier methods, OverlapNet learns end-to-end from multiple LiDAR modalities and estimates matching quality through overlap.
III. OUR APPROACH
The approach defines LiDAR overlap on spherical range-image projections and uses it as a transformation-independent signal for loop closure detection. Correct-pose overlap is used for training, while online detection avoids unreliable drifted poses.
- Loop Closure Detection: Overlap thresholds can identify scans at the same place without knowing their correct relative pose.This makes overlap useful when accumulated SLAM drift makes pose-based overlap calculation unreliable online.
- Spherical Projection: LiDAR point clouds are spherical-projected into vertex maps, with each pixel representing the nearest 3D point.The projection maps 3D points to image coordinates using sensor spherical coordinates.
- The Concept of Overlap: For two scans, one vertex map is reprojected into the other scan’s frame before corresponding range differences are computed.
- The Concept of Overlap: Overlap measures the percentage of corresponding valid range-image pixels whose reprojected ranges differ by at most ϵ.Only pixels with valid readings in both range images contribute.
C. Overlap Network Architecture
OverlapNet combines multiple LiDAR-derived cues in a spherical-projection architecture to estimate scan overlap and relative yaw for loop closing.
- The network input combines range, normal, intensity, and semantic class probability information generated from each LiDAR scan.Semantic probabilities from RangeNet++ are compressed before inclusion in the input tensor.
- Direct overlap computation without accurate relative poses produces high values for many scan pairs, whereas OverlapNet predicts overlap for loop-closure selection.The figure contrasts pose-dependent Eq. (3) estimates with learned predictions.
- OverlapNet uses two weight-sharing legs and two heads that operate on their paired feature volumes.The heads produce overlap and relative yaw estimates between two scans.
1) Legs:
The architecture extracts compact, yaw-aware feature volumes and uses a delta head to encode pairwise differences for overlap estimation.
- Legs:: Each weight-sharing leg is an 11-layer fully convolutional network producing a 1×360×128 feature volume.The lightweight feature representation preserves the cyclic horizontal structure of spherical LiDAR projections.
- Legs:: Because yaw rotation causes a cyclic column shift, the single feature-volume row represents relative yaw through horizontal displacement.This design exploits the geometry of the range-image projection.
- 2) Delta Head:: The delta head estimates overlap from the two feature volumes using a delta layer, three convolutional layers, and one fully connected layer.The delta layer computes absolute differences across all pixel pairs after tiling and transposing one feature volume.
- 2) Delta Head:: Different overlaps produce different patterns in the delta-layer output that subsequent layers use for overlap prediction.The representation captures latent differences between the paired scans.
3) Correlation Head:
The correlation head estimates relative yaw from cyclic feature matching, while joint losses train overlap regression and yaw classification with a restricted yaw-training condition.
- 3) Correlation Head:: The correlation head cross-correlates cyclic feature volumes and produces a one-dimensional output of size 360 for yaw estimation.One feature volume is horizontally padded and the other is shifted as a kernel; the argmax gives the angle in degrees.
- D. Loss Functions: The total training loss combines overlap loss and yaw-angle loss using weight α.The overlap output is trained with weighted absolute difference, while the yaw head uses binary cross-entropy.
- D. Loss Functions: Yaw estimation is formulated as binary classification over output entries rather than direct regression because argmax is nondifferentiable.Binary cross-entropy is used for this classification objective.
- D. Loss Functions: Yaw-angle training is limited to scan pairs with overlap larger than 30% because this minimum was needed for correct ICP pose estimates.The threshold was also experimentally validated in the paper.
E. SLAM Pipeline
OverlapNet replaces SuMa’s nearest-frame loop-closure heuristic within a surfel-based SLAM pipeline whose scan alignment uses projective ICP and iterative Gauss–Newton optimization.
- E. SLAM Pipeline: OverlapNet is integrated into SuMa by replacing its original heuristic loop-closure detection module.SuMa is the surfel-based mapping system used as the SLAM pipeline.
- E. SLAM Pipeline: SuMa estimates scan transformations with projective ICP against rendered map views and incrementally chains pose updates.The optimization uses point-to-plane error over projective vertex associations.
- E. SLAM Pipeline: Gauss–Newton optimization determines pose increments by iteratively solving the weighted linearized system.The Jacobian describes residual sensitivity to the six-dimensional increment.
- E. SLAM Pipeline: SuMa’s nearest-frame heuristic works for small loops but fails in areas containing only a few large loops.The heuristic depends on the current pose estimate and can select incorrect candidates when drift causes large displacements.
F. Covariance Propagation for Geometric Verification
The method propagates pose uncertainty through incremental estimation and uses a scaled probabilistic distance to restrict loop-closure search, while also supporting place recognition without prior pose information.
- The noisy relative pose is represented by a mean and covariance, with covariance estimated from robustized point-to-plane residuals.The covariance uses Huber correction factors, squared residuals, correspondence count, and a six-dimensional pose transformation.
- Covariance propagation updates the mean and covariance during incremental pose estimation.The update uses the Jacobian of the relative transformation.
- Lie algebra expresses poses as six-dimensional vectors so the Mahalanobis distance can compare pose estimates probabilistically.The transformation is mapped using ξ = log T.
- The scaled distance restricts the search space according to pose uncertainty, reducing computation while permitting place recognition without prior information.Unlike a fixed search radius, the uncertainty-aware strategy adapts the search region automatically.
IV. EXPERIMENTAL EVALUATION
The evaluation tests loop-closure detection on KITTI and an unseen Ford campus dataset under defined search and overlap criteria, comparing OverlapNet with established methods and internal variants.
- The experiments assess overlap and yaw prediction, loop-closure candidate detection, SLAM mapping improvement, and detection without prior pose information.These claims are evaluated through benchmark comparisons, mapping experiments, and odometry-free place recognition.
- KITTI sequence 00 is used for evaluation, sequences 03–10 for training, and sequence 02 for validation.The scans come from urban areas around Karlsruhe and were recorded with a Velodyne HDL-64E.
- The Ford campus test uses sequence 00 and a model never trained on that dataset, measuring generalization across sensor setup and environment.The Ford data were recorded in Michigan with a different Velodyne HDL-64E version.
- A 30% overlap threshold defines true positives, allowing evaluation of challenging revisits with low scan overlap.The paper notes that ICP can recover correct poses at approximately this overlap.
- OverlapNet is compared with M2DP, Histogram, SuMa, and a covariance-based top-10 variant using precision-recall curves.SuMa contributes a single precision-recall point because it always selects the nearest frame.
- OverlapNet outperforms other methods on KITTI, is similar on Ford campus, and surpasses all baselines on both datasets when prior information is integrated.On Ford campus, evaluation uses the geometry-only model because intensity, normals, and semantic cues are unavailable.
- The AllChannel two-head model substantially outperforms the MLPOnly variant, indicating that the simpler head is insufficient for strong performance.The comparison uses AUC and F1 scores.
B. Qualitative Results
Integrating OverlapNet improves qualitative SLAM behavior by producing more accurate loop-closure candidates and enabling additional challenging revisits, while odometry-free detection is strongest with few candidates.
- After integration, the SLAM system produces more accurate odometry on KITTI sequence 02 than SuMa.The comparison visualizes 3D translation error relative to ground truth.
- OverlapNet enables the system to find more loops when the vehicle revisits an intersection from the opposite direction.The paper identifies this as a challenging low-overlap situation.
- Without prior pose information, OverlapNet outperforms all baselines with a small number of candidates and matches them at larger candidate counts.OREOS and LocNet++ achieve slightly higher recall when more candidates are considered.
D. Yaw Estimation
OverlapNet estimates relative yaw accurately, especially at high overlap, and benefits from correlation-based processing and complementary input modalities.
- OverlapNet outperforms the other methods in mean yaw error and standard deviation on KITTI sequence 00.The evaluation summarizes yaw-angle errors without ICP.
- OverlapNet and OREOS always provide yaw estimates, whereas the RANSAC-based method sometimes fails.Both methods are designed to estimate relative yaw for every scan pair.
- The correlation head uses shifts in the range projection to represent orientation, and joint overlap-yaw training further improves results.The paper attributes the advantage over OREOS’s multilayer perceptron partly to this representation.
- For overlaps above 90%, the relative yaw estimate has an average error of about 1 degree.Yaw estimation improves as real overlap increases.
- Depth alone gives reasonable overlap prediction but weak yaw estimation, while adding normals supports good performance on both tasks.Using more input modalities makes the method more robust.
- Semantic information can distinguish scans in symmetrical environments, but geometric information alone can also achieve good performance.Semantic inputs additionally increase computation time.
F. Using OverlapNet Predictions as Initial Guesses for ICP
OverlapNet predictions are used to initialize ICP for 3D LiDAR scan matching, with yaw estimates improving registration initialization. The method is designed for online SLAM operation by restricting candidate evaluation to a pose-uncertainty search space.
- OverlapNet predicts relative yaw angles that initialize ICP registration for 3D LiDAR scans.The registration error is measured as Euclidean distance between estimated and ground-truth translations.
- Fig. 9 relates scan overlap to ICP registration error with and without OverlapNet initial guesses.The experiment evaluates whether predicted overlap and yaw provide useful initialization across different overlap levels.
- 75 ms preprocessing, 6 ms leg-feature extraction, and 27 ms head matching were required per KITTI frame on the reported system.Worst-case head matching over all candidates took 630 ms.
- The approach is intended to support online operation because Mahalanobis-distance search usually limits evaluation to a small candidate set.This restriction applies during real SLAM operation rather than evaluating the entire search space.