BenchRNR: Roadside LiDAR Localization Benchmark

• The paper introduces BenchRNR, a benchmark that compares repetitive and non-repetitive LiDAR scanning for centimeter-level vehicle localization.
• It details real-world sensor setups, synchronization protocols, and calibration techniques to derive accurate ground truth in low-speed parking scenarios.
• Baseline evaluations reveal that non-repetitive LiDAR, when integrated with robust template registration, delivers performance comparable to costly repetitive sensors.

BenchRNR is most directly associated with “Bench-RNR: Dataset for Benchmarking Repetitive and Non-repetitive Scanning LiDAR for Infrastructure-based Vehicle Localization,” a real-world roadside LiDAR benchmark introduced to compare repetitive scanning LiDAR and non-repetitive scanning LiDAR under the same infrastructure-based vehicle-localization conditions (Zhao et al., 19 Sep 2025). The benchmark addresses a practical deployment question in intelligent transportation systems: when a roadside sensor is used to localize vehicles with centimeter-level accuracy for cloud-controlled vehicles, should one deploy a conventional rotating multi-line LiDAR or a non-repetitive LiDAR such as Livox. In its hyphenated form, Bench-RNR denotes the roadside localization dataset; an unrelated 2026 paper also uses “BenchRNR” to denote the NHANES Accelerometry Cardiometabolic Benchmark for digital biomarkers (Felizzi, 29 Jun 2026).

1. Research problem and benchmark motivation

Bench-RNR was created to support infrastructure-based vehicle localization rather than generic roadside perception. The paper situates the problem in roadside localization for cloud-controlled vehicles, where a fixed roadside sensor can serve multiple vehicles simultaneously and can avoid the cost of placing expensive sensors on every car (Zhao et al., 19 Sep 2025).

The motivation is explicitly comparative. Most existing studies rely on repetitive scanning LiDARs, whereas non-repetitive scanning LiDARs are described as eliminating blind zones and being more cost-effective. At the same time, the two scanning patterns differ substantially in coverage, point distribution, and cost. Bench-RNR was therefore designed so that their localization performance can be directly compared under the same real-world conditions.

The benchmark also responds to a dataset gap. Prior roadside datasets such as DAIR-V2X-I, TUMTraf Intersection, and V2X-Real are described as being primarily aimed at 3D detection and perception rather than localization, and as typically lacking the centimeter-level ground truth needed to validate vehicle-localization methods. Bench-RNR addresses that gap by collecting synchronized roadside and vehicle-side measurements and by using a high-precision vehicle-end GNSS/IMU system to define localization ground truth.

2. Dataset composition and scene characteristics

Bench-RNR contains 5,445 frames of point clouds across 8 vehicle trajectory sequences collected in an open-air parking lot at Shanghai Jiao Tong University (Zhao et al., 19 Sep 2025). The site includes 11 distinct parking spots, and the trajectories include both parking and driving sequences with diverse route types.

The dataset is intentionally realistic but controlled. The paper emphasizes that vehicle motions are typical of parking scenarios, so the benchmark is dominated by low-speed maneuvers, with an average speed of 4.8 km/h. Vehicle orientations are also analyzed: heading angles are concentrated around horizontal and vertical directions because of the parking-lot geometry, though other directions are represented as well.

This composition makes the dataset particularly suited to infrastructure-localization studies in settings where low-speed precision matters. The sequences are not presented as a general-purpose urban-driving corpus; rather, they are a controlled benchmark for analyzing how scanning patterns behave across diverse vehicle poses and scene configurations within a roadside localization regime.

3. Sensor suite and the repetitive versus non-repetitive distinction

The benchmark’s central comparison is implemented through a mixed roadside sensor suite consisting of one repetitive scanning LiDAR, two non-repetitive scanning LiDARs, and two cameras (Zhao et al., 19 Sep 2025). The repetitive scanner is a Hesai OT128 128-line LiDAR mounted 19.25 m above the scene and tilted downward to cover about 160 degrees of the environment. The non-repetitive scanners are Livox Avia sensors with a 70.4° × 77.2° field of view.

The paper draws a strong conceptual distinction between the two scanning patterns. Repetitive scanning LiDAR traces the same scan pattern over time, producing a more regular and often denser point cloud at any instant. However, when such a sensor is installed in a fixed roadside position, the discrete scan lines create fixed blind zones between lines. Non-repetitive scanning LiDAR, by contrast, follows a non-repeating scan trajectory that progressively fills the field of view over time. Individual frames are sparse and unevenly distributed, but temporal accumulation can cover the scene more completely and can eliminate fixed blind zones.

The benchmark also encodes a deployment-oriented cost-performance tradeoff. The paper gives representative values of 3.46M pts/s for a Hesai OT128 and 240k pts/s for a Livox Avia, with a price difference of roughly \$140,000** versus **\$1,400. The core limitation of non-repetitive scanning within the benchmark is that a single frame often contains fewer points on the target vehicle, which makes shape-based localization harder unless the localization method is robust to incompleteness.

4. Data format, synchronization, calibration, and ground truth

Bench-RNR is released in ROS bag format, with repetitive LiDAR stored as standard point clouds and non-repetitive Livox data stored in Livox CustomMsg format so that per-point timestamps are preserved (Zhao et al., 19 Sep 2025). The preservation of per-point timestamps is important because it supports correction of intra-frame distortions caused by moving objects and enables temporal processing.

Temporal alignment is performed over a shared local area network using NTP. The roadside sensors are triggered at 10 Hz, and the vehicle-mounted GNSS/IMU is recorded at 100 Hz. This synchronization regime is part of the benchmark design because the comparison between scanning patterns is only meaningful when the roadside and vehicle-side measurements are precisely time-aligned.

Extrinsic calibration aligns all LiDAR coordinate systems to a common world frame. The paper reports a two-stage process: RANSAC is used to estimate the ground plane, after which overlapping point clouds are manually aligned to recover translation and yaw. Ground-truth vehicle localization is then derived by extracting the vehicle point cloud from each frame, matching it to a high-precision vehicle template to estimate the vehicle center, and optimizing the transformation from GNSS UTM coordinates to the site world coordinate system. Vehicle orientation ground truth comes from IMU integration.

5. Baseline localization pipeline and evaluation protocol

To benchmark infrastructure-based localization, the authors implement a pipeline that first segments the target vehicle point cloud using a rule-based method comprising background modeling, background filtering, clustering, and greedy multi-frame association (Zhao et al., 19 Sep 2025). The segmented vehicle cloud is then processed by four baseline localization methods.

Method	Non-repetitive scanning	Repetitive scanning (OT128)
Seg+obb	16.67 cm, 3.69°	14.23 cm, 4.77°
Seg+convex	15.56 cm, 1.35°	13.06 cm, 1.60°
PV-RCNN	15.79 cm, 2.84°	16.06 cm, 2.70°
Register-Loc	6.87 cm, 1.03°	6.84 cm, 0.88°

The four baselines are defined as follows. Seg+obb uses PCA to estimate the principal direction and then computes an oriented bounding box. Seg+convex applies a convex-hull-based pose estimator. Register-Loc registers the segmented cloud to a high-precision vehicle template and infers the box from template alignment. PV-RCNN is a learned point-cloud detector retrained on roadside data from DAIR-V2X-I using OpenPCDet, because a model trained directly on vehicle-side data did not transfer well.

The evaluation reports localization error and heading estimation error. In the paper’s reported results, localization values are given in centimeters and heading values in degrees. The benchmark therefore tests both translational and rotational accuracy under matched roadside sensing conditions.

6. Findings, interpretation, and benchmark significance

The central empirical result is that the non-repetitive LiDAR achieves localization performance comparable to 128-line and pseudo-64-line repetitive LiDAR across all four evaluated methods (Zhao et al., 19 Sep 2025). In some cases it slightly outperforms the 64-line version, especially with Register-Loc and PV-RCNN. In particular, Register-Loc reaches about 6.87 cm mean localization error with non-repetitive scanning, which the authors argue is already sufficient for safe cloud-controlled vehicle operation.

The paper’s recommendation is deliberately nuanced. Repetitive scanning remains advantageous when the localization method relies on a more complete or more uniformly distributed instantaneous vehicle shape. Methods such as simple PCA bounding boxes or convex-hull fitting generally perform better or more stably under repetitive scanning because the point distribution on the vehicle is denser and more regular. Non-repetitive scanning, by contrast, produces sparse and uneven observations, which degrades methods that assume a full vehicle contour.

At the same time, the benchmark shows that non-repetitive scanning is a viable lower-cost choice when paired with a localization method that incorporates prior knowledge of the complete vehicle shape. The authors therefore conclude that non-repetitive roadside LiDAR is especially suitable when used with robust model-based localization such as template registration, which compensates for sparsity and uneven sampling.

Bench-RNR and its source code are publicly available at https://github.com/sjtu-cyberc3/BenchRNR (Zhao et al., 19 Sep 2025). The release is positioned as a community resource for infrastructure-based perception and vehicle localization, with future plans to expand to more roadside scenarios and to annotate additional tasks such as object detection. Within the roadside LiDAR literature, its significance lies not only in the introduction of a new dataset, but in the provision of comparative evidence about when repetitive and non-repetitive scanning patterns are preferable for infrastructure-based vehicle localization.