Source-linked AI summary
ERASOR: Egocentric Ratio of Pseudo Occupancy-based Dynamic Object Removal for Static 3D Point Cloud Map Building
Hyungtae Lim, Sungwon Hwang, Hyun Myung
TL;DR
Dynamic objects in sequential urban LiDAR accumulations leave map traces that impede mobile-robot localization and navigation. ERASOR uses pseudo occupancy, Scan Ratio Test, and Region-wise Ground Plane Fitting to remove such points without visibility tracing. On SemanticKITTI, it shows promising performance against state-of-the-art ray tracing-based and visibility-based methods.
Problem
Sequential urban LiDAR accumulations preserve dynamic-object traces that act as obstacles and impede mobile-robot localization and navigation.
Method
ERASOR uses vertical pseudo-occupancy columns, Scan Ratio Test to identify candidate bins, and Region-wise Ground Plane Fitting to distinguish static from dynamic points.
Results
On SemanticKITTI, ERASOR shows promising performance against state-of-the-art methods, removing dynamic points with the least loss of static points and operating at least ten times faster.
Takeaways & Limitations
The resulting static map contains few dynamic objects and helps improve mobile-robot navigation and localization tasks.
Abstract
from arXiv · showhide
Scan data of urban environments often include representations of dynamic objects, such as vehicles, pedestrians, and so forth. However, when it comes to constructing a 3D point cloud map with sequential accumulations of the scan data, the dynamic objects often leave unwanted traces in the map. These traces of dynamic objects act as obstacles and thus impede mobile vehicles from achieving good localization and navigation performances. To tackle the problem, this paper presents a novel static map building method called ERASOR, Egocentric RAtio of pSeudo Occupancy-based dynamic object Removal, which is fast and robust to motion ambiguity. Our approach directs its attention to the nature of most dynamic objects in urban environments being inevitably in contact with the ground. Accordingly, we propose the novel concept called pseudo occupancy to express the occupancy of unit space and then discriminate spaces of varying occupancy. Finally, Region-wise Ground Plane Fitting (R-GPF) is adopted to distinguish static points from dynamic points within the candidate bins that potentially contain dynamic points. As experimentally verified on SemanticKITTI, our proposed method yields promising performance against state-of-the-art methods overcoming the limitations of existing ray tracing-based and visibility-based methods.
I. INTRODUCTION AND RELATED WORKS
Sequential LiDAR accumulations can preserve dynamic objects as map traces that obstruct localization and navigation. ERASOR addresses this through pseudo occupancy, ground-contact cues, and region-wise fitting, targeting limitations of prior approaches.
- Sequentially accumulated urban scans can leave vehicle and pedestrian traces that act as obstacles for mobile-robot localization and navigation.
- Existing methods include segmentation-based, ray tracing-based, and visibility-based approaches for rejecting dynamic objects from generated maps.
- Segmentation-based methods depend heavily on supervised labels, making them vulnerable to labeling errors and unlabeled dynamic-object classes.
- Ray tracing counts scan hits and misses in voxel spaces, but its computational expense motivated lower-cost visibility-based alternatives.
- Visibility-based methods face incidence-angle ambiguity and occlusion, preventing reliable removal of large dynamic objects when static points are not observed behind them.
- ERASOR is a visibility-free static-map method that represents points in vertical pseudo-occupancy columns, uses Scan Ratio Test to find candidate bins, and applies R-GPF to retrieve static points.The method is presented as fast and robust, with quantitative validation against state-of-the-art methods on SemanticKITTI using Preservation Rate and Rejection Rate.
- ERASOR reportedly removes dynamic points with the least loss of static points and operates at least ten times faster than state-of-the-art methods.
II. STATIC MAP BUILDING VIA ERASOR
ERASOR builds a static map by comparing query and map point clouds in a visibility-free representation, identifying potentially dynamic spaces, and refining them using ground structure. Its scope is instance-level dynamic-object rejection, excluding large-scale urban changes.
- The approach is visibility-free and organizes points into vertical columns represented through pseudo occupancy.The paper presents ERASOR as a response to limitations of ray tracing-based and visibility-based methods.
- The method focuses on instance-level dynamic-object rejection and excludes redevelopment or restoration of buildings from its scope.
- ERASOR constructs a static map by analyzing discrepancies between a query cloud and a prior map containing measured dynamic points.The estimated dynamic points are determined from discrepancies between query and map clouds.
- The method assumes most urban dynamic objects, including vehicles and pedestrians, are in contact with the ground.This assumption yields four unit-space cases, with the first categorized as potentially dynamic and the third and fourth as definitely static.
- ERASOR detects potentially dynamic unit spaces and retrieves regions containing dynamic points before removing those dynamic parts from the map cloud.The second ground-object case is neglected because the objective is to refine the existing map cloud.
B. Volume of Interest
The Volume of Interest restricts processing to a bounded radial and vertical region around the query cloud, capturing the expected height range of target dynamic objects while reducing computation.
- The Volume of Interest partitions relevant point space using radial and vertical bounds around the query cloud.It includes points satisfying ρ_k < Lmax and hmin < z_k < hmax.
- The VOI parameters are set to hmax = 3.0m, hmin = −1.0m, and Lmax = 80.0m.The lower height bound covers sloped areas and uncertainty in detected ground height.
- The selected height range is intended to cover dynamic objects such as vehicles and pedestrians.
- The map subregion is extracted by neighbor search near the query frame before transformation, avoiding transformation of the full map.This reduces computational cost because the map may contain more than a million points.
C. Region-wise Pseudo Occupancy Descriptor
R-POD encodes the VOI as an egocentric arrangement of rings and sectors, using vertical boundary differences to represent pseudo occupancy in each unit space.
- R-POD is an egocentric spatial occupancy descriptor designed for region-wise dynamic-object removal.It encapsulates the VOIs and is motivated by Cells of Interest and Scan Context.
- R-POD avoids relying on absolute maximum height alone when comparing scan-to-map relationships.The paper contrasts this design with Scan Context’s maximum-height encoding.
- R-POD divides the VOI into regular azimuthal and radial intervals called sectors and rings.The descriptor uses Nr rings and Nθ sectors.
- Each vertical bin expresses occupancy through the boundary difference of its vertical information, called pseudo occupancy.
- For each bin, the descriptor groups cloud points satisfying the bin’s spatial condition and computes pseudo occupancy from their vertical coordinates.The vertical-coordinate set is denoted Z(i,j),t, with supremum and infimum defining the occupancy boundaries.
D. Scan Ratio Test
The Scan Ratio Test compares pseudo occupancy between corresponding query and map bins to identify potentially dynamic regions. Dynamic-containing bins tend to have low ratios, while static-only bins have ratios near one.
- SRT compares corresponding query and map R-POD bins to detect discrepancies in pseudo occupancy.It applies a ratio-test principle motivated by Lowe’s Ratio Test in SIFT.
- A scan ratio near 1 indicates unchanged static bins, whereas a much smaller ratio indicates a potentially dynamic case.The lower ratio is associated with a dynamic object on the ground.
- Bins with a scan ratio below the threshold are selected as potentially dynamic bins.The threshold is empirically set to 0.2.
- SemanticKITTI distributions show that bins containing dynamic points tend to have small scan ratios, while static-only bins tend to have ratios close to 1.Figure 4 presents the ratio PDF and the histogram of actual cloud-point counts for corresponding bins.
- SRT is skipped when either the map or query bin contains few points.
E. Region-wise Ground Plane Fitting
Region-wise Ground Plane Fitting (R-GPF) retrieves likely static points from candidate bins by fitting local ground planes and retaining points below them. Its bin-wise design accommodates nonplanar urban terrain, while iterative PCA estimates each plane and the resulting points are reverted into the map.
- R-GPF selects candidate bins that potentially contain dynamic points, exploiting the observation that many such bins contain only ground and dynamic points.
- Bin-wise plane fitting is preferred because local urban grounds with curbs or slopes may not form one perfect plane across the whole map.Each small bin is assumed to be locally planar.
- The method initializes ground inliers from the lowest-height seed points, using their mean height and a height margin.
- PCA estimates the ground normal as the eigenvector with the smallest covariance eigenvalue, then derives the plane coefficients.
- After three iterations, the final ground inliers are reverted into the map, while points rejected by the procedure are treated as dynamic points.
- R-GPF represents dynamic points in red and reverted ground points in green in its procedural visualization.
A. Dataset
The evaluation uses SemanticKITTI point-wise annotations and a selected benchmark of five dynamic-object-heavy frame ranges. These ranges cover countryside, highway, and intersection environments, with maps constructed using poses from SuMa.
- SemanticKITTI point-wise annotations provide ground truth, with selected annotated classes labeled as dynamic points to remove.The selected class identifiers are 252, 253, 254, 255, 256, 257, and 259.
- The quantitative evaluation manually selects the top-five time frames containing the largest number of dynamic-object appearances.This selection makes dynamic traces more distinguishable than evaluation over frames containing few or no dynamic points.
- The benchmark comprises frame ranges 00, 01, 02, 05, and 07, spanning countryside, highway, and intersection environments.The ranges are 4,390–4,530; 150–250; 860–950; 2,350–2,670; and 630–820, respectively.
- Maps are constructed at regular intervals using SuMa poses, which contain inherent uncertainty.
B. Error Metrics
The paper introduces Rejection Rate (RR) and Preservation Rate (PR) for static-map evaluation, arguing that conventional precision-recall is insensitive and voxelization-variant. The metrics compare preserved static or dynamic points after identical voxelization and use nearest-point preservation checks.
- Rejection Rate (RR) and Preservation Rate (PR) are proposed as static-map-oriented quantitative metrics.
- Precision-recall can remain too low and insensitive because false positives from erased static points greatly outnumber true positives from erased dynamic points.
- Precision-recall is voxelization-variant because different algorithms use different voxel sizes, changing their false-positive and true-positive counts.
- PR is the number of preserved static points divided by the total static points on the raw map, while RR is one minus preserved dynamic points divided by total dynamic points.
- PR and RR are calculated voxel-wise after applying identical voxelization with voxel size 0.2 to baseline static maps.
- The preservation function compares each voxelized raw-map point with its nearest retrieved-map point using inter-voxel inclusion and matching static-or-dynamic labels.Inter Voxel Inclusion returns true when two points occupy the same voxel; the label function returns true when both are static or both dynamic.
IV. RESULTS AND DISCUSSION
ERASOR’s results show a preservation–rejection trade-off controlled by the ground threshold, while R-GPF and scan-ratio-based removal improve static-map quality against prior methods.
- A. Impact of the Ground Threshold: Increasing τg improved Preservation Rate but reduced Rejection Rate because more near-ground dynamic points were classified as ground.Preservation Rate increased most sharply from 0.05 to 0.15, while Rejection Rate decreased most sharply from 0.15 to 0.25.
- B. Comparison of R-GPF with Ground Plane Fitting: R-GPF achieved higher static point retrieval performance than GPF across varied urban environments.Its largest advantage occurred on sequence 00, where curbs created non-planar floors.
- C. Comparison with the State-of-the-Art Methods: ERASOR removed dynamic traces that remained for state-of-the-art methods on bus tops in sequences 05 and 07.Its scan-ratio check can remove points outside the valid range for ray tracing or visibility checks.
- C. Comparison with the State-of-the-Art Methods: OctoMap achieved strong dynamic-point rejection but falsely removed many static points, producing the lowest Preservation Rate.OctoMap and Peopleremover lost relatively more static points than ERASOR, consistent with motion ambiguity.
- C. Comparison with the State-of-the-Art Methods: ERASOR produced promising Preservation Rate and F1 scores in all environments except sequence 01.The lower Preservation Rate in sequence 01 was attributed to removing vegetation far from the highway after partial observation.
D. Algorithm speed
ERASOR was evaluated against state-of-the-art methods and achieved the fastest one-iteration runtime, with its bin-wise scan-ratio procedure enabling one-shot dynamic-point removal.
- D. Algorithm speed: ERASOR showed the fastest performance for one iteration and was at least ten times faster than the other methods.Removert requires multiple spherical-image projections, whereas ERASOR removes dynamic points in one shot through a bin-wise scan-ratio operation.
- D. Algorithm speed: ERASOR was quantitatively compared with OctoMap, Peopleremover, and Removert on SemanticKITTI.The comparison used reported Preservation Rate and Rejection Rate metrics for static-map building.
- D. Algorithm speed: The study tested ERASOR quantitatively and qualitatively on SemanticKITTI.The conclusion reports promising results across various environments and few dynamic objects in the resulting static maps.