RoundaboutHD: HD Map-Aware Tracking Benchmark

• RoundaboutHD is a suite of high-definition roundabout resources that integrates map-aware perception, multi-camera vehicle tracking, and HD-map registration in varied applications.
• It comprises both a formal 2025 multi-camera benchmark with 4K recordings and a continuous roadside deployment in Ann Arbor using robust calibration and trajectory extraction.
• The ecosystem combines integrated sensing, advanced computer vision pipelines, and procedural map generation to tackle localization, tracking, and autonomous driving challenges.

RoundaboutHD denotes a family of high-definition roundabout resources centered on map-aware perception, tracking, and autonomous-driving research. In the literature, the name is used explicitly for a high-resolution, real-world benchmark for multi-camera vehicle tracking, while related work uses it functionally for an HD-map–registered roadside perception deployment in Ann Arbor and, more loosely, for HD-mapped roundabout corpora such as openDD (Lin et al., 11 Jul 2025, Zou et al., 2022, Breuer et al., 2020). The common thread is not a single canonical dataset format, but the combination of roundabout geometry, trajectory-centric annotations or estimates, and a representation rich enough to support localization, multi-camera association, or map-conditioned planning.

1. Terminology and referential scope

A recurring source of confusion is that “RoundaboutHD” does not denote a single universally fixed artifact across the literature. In one line of work it is the proper name of a benchmark dataset. In another, it is a functional label for an always-on roadside deployment and its HD-map–registered data stream. In a third, it serves as a descriptive shorthand for any roundabout dataset that couples trajectories with HD map topology.

Referent	Status in the literature	Core characteristics
Ann Arbor roadside deployment	Functional use; the term is not explicitly used in the paper	Two-lane roundabout, HD-map–anchored homography, 7×24 monitoring, sub-meter trajectory extraction
openDD	Descriptive use; no separate dataset named “RoundaboutHD”	Seven roundabouts, drone trajectories, HD maps via shapefiles and XML
"RoundaboutHD: High-Resolution Real-World Urban Environment Benchmark for Multi-Camera Vehicle Tracking"	Explicit dataset name	Four non-overlapping 4K cameras, 15 fps, 40 minutes, 512 identities

The Ann Arbor work states that the term is not explicitly used in the paper’s text, but identifies a named, continuously operating roadside camera system at Ellsworth Rd. and State St. built around a landmark-registered homography to an HD geographic frame (Zou et al., 2022). The openDD paper likewise does not introduce a separate dataset named “RoundaboutHD”; rather, openDD itself provides HD roundabout maps and trajectory data at scale (Breuer et al., 2020). By contrast, the 2025 benchmark formally introduces “RoundaboutHD” as a comprehensive, high-resolution multi-camera vehicle tracking benchmark dataset (Lin et al., 11 Jul 2025).

2. Infrastructure-assisted roadside perception in Ann Arbor

In the Ann Arbor usage, RoundaboutHD refers to a full-scale infrastructure-assisted deployment at the two-lane roundabout located at Ellsworth Rd. and State St. in Ann Arbor, MI. Its stated objectives are infrastructure-assisted autonomous driving, traffic monitoring across all lanes and conflict points, hazardous driving warnings, and high-precision vehicle trajectory extraction over the roughly $10{,}000\ \mathrm{m}^2$ roundabout area (Zou et al., 2022).

The sensing layout comprises eight cameras mounted around the roundabout’s periphery: four $360^\circ$ fisheye cameras and four long-range pinhole thermal cameras, each positioned at the four corners of the roundabout. The overlap pattern is asymmetric by design. Fisheyes provide near-field panoramic coverage within about $25\ \mathrm{m}$ of each pole, while thermals extend coverage outward along entries and exits to within approximately $200\ \mathrm{m}$ in their field of view. The framework is designed to operate on either optical or thermal modalities and supports both pinhole and fisheye lenses without joint calibration to other sensors (Zou et al., 2022).

Its calibration and localization stack is purely vision-based and is deliberately decoupled from detection. For each camera, the authors manually select $5$–$20$ ground landmarks and annotate both pixel coordinates and real-world longitude/latitude using Google Maps. The road surface is modeled as piecewise planar segments, and for each segment a planar projective transform $H$ is estimated via least squares with RANSAC consensus across landmark pairs. Geographic lookup masks are then precomputed so that detected bottom-center pixels can be lifted directly into the HD geographic frame:

$$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$

For fisheye cameras, the method first assumes a generic radially symmetric distortion model

$$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$

and estimates the intrinsic matrix $K$ and distortion coefficients before homography estimation on the undistorted domain (Zou et al., 2022).

The perception pipeline is modular and contains four decoupled modules: object detection, object localization, multi-camera information fusion, and object tracking. Detection uses a single-stage, center-aware architecture with a MobileNet-v2 encoder, an FPN-style decoder, and four output heads for bottom-center prediction, 3D size, yaw pose, and category recognition. The multi-task loss is

$360^\circ$0

with $360^\circ$1. No depth offset or 3D labels are needed; training uses only 2D labels consisting of the bottom rectangle of each vehicle (Zou et al., 2022).

Tracking is performed in world coordinates rather than image coordinates, which the paper emphasizes as a robustness mechanism for fisheye views. The Kalman filter state is

$360^\circ$2

and the tracker builds on SORT with Hungarian assignment and a maximum age of $360^\circ$3 frames (Zou et al., 2022).

The reported performance establishes the deployment as a real-time, edge-executable HD roundabout sensing system. Detection AP at VOC IoU $360^\circ$4 is $360^\circ$5 on AA-Fisheye and $360^\circ$6 on AA-Thermal. Average fisheye in-ROI calibration error is $360^\circ$7, and thermal in-ROI calibration error is $360^\circ$8. Using an instrumented Hybrid Lincoln MKZ with RTK GPS and IMU, the average trajectory localization error is $360^\circ$9 for fisheye cameras in-ROI and $25\ \mathrm{m}$0 for thermal cameras in-ROI. Fusion of all four fisheyes reduces the average error across the whole roundabout from $25\ \mathrm{m}$1 to $25\ \mathrm{m}$2. On a Jetson AGX Xavier, single-stream half-precision inference reaches $25\ \mathrm{m}$3 with $25\ \mathrm{m}$4 end-to-end delay, and the abstract states an all-component end-to-end delay of less than $25\ \mathrm{m}$5 (Zou et al., 2022).

3. The explicit 2025 multi-camera vehicle tracking benchmark

The explicit benchmark named RoundaboutHD is a high-resolution, real-world dataset designed for multi-camera vehicle tracking in an urban roundabout. It was created to address limitations of publicly available datasets, specifically overly simplistic scenarios, low-resolution footage, and insufficiently diverse conditions. The dataset consists of four non-overlapping surveillance cameras positioned around a roundabout in a US town, with a maximum inter-camera distance of approximately $25\ \mathrm{m}$6. All cameras record at $25\ \mathrm{m}$7 resolution $25\ \mathrm{m}$8 and $25\ \mathrm{m}$9, with $200\ \mathrm{m}$0 minutes per camera and $200\ \mathrm{m}$1 minutes total. The cameras start recording simultaneously to support temporally aligned multi-camera tracking (Lin et al., 11 Jul 2025).

The core annotation set contains $200\ \mathrm{m}$2 unique vehicle identities and $200\ \mathrm{m}$3 validated bounding boxes. Per-camera vehicle counts are $200\ \mathrm{m}$4, $200\ \mathrm{m}$5, $200\ \mathrm{m}$6, and $200\ \mathrm{m}$7 for cam01 through cam04, and $200\ \mathrm{m}$8 vehicles appear in at least two cameras. The benchmark provides manually verified bounding boxes, per-camera IDs, cross-camera IDs, vehicle attributes, and camera modeling/geometry parameters. The attribute inventory includes color, type $200\ \mathrm{m}$9, make, and model. The most frequent makes are Chevrolet, Ford, Honda, Toyota, and Nissan; common models include Chevrolet Silverado, Ford Transit, Honda CR‑V, Chevrolet Equinox, and Toyota Camry (Lin et al., 11 Jul 2025).

The dataset is explicitly partitioned into multiple task-specific subsets. The object-detection subset contains $5$0 boxes with type labels. The single-camera tracking subset contains $5$1 annotated trajectories totaling $5$2 boxes, with the trajectory count exceeding unique IDs because long-term occlusions sometimes cause ID splits. The image-based vehicle ReID subset contains $5$3 cropped images, of which the test set uses $5$4 gallery images and $5$5 query images from $5$6 identities, while the train set contains $5$7 images from $5$8 identities. ReID crops follow a minimum bounding-box resolution of $5$9 pixels; gallery images are sampled every $20$0 frames and query images every $20$1 frames, prioritizing high-quality crops (Lin et al., 11 Jul 2025).

Its annotation workflow is semi-automatic. Detection is initialized with YOLOv12x pretrained on COCO using confidence threshold $20$2 and area filter $20$3 pixels. Deep features are extracted with ResNet-101, tracklets are generated with SMILEtrack, and cross-camera association is finalized by human annotators using visual attributes and spatiotemporal consistency. Camera modeling follows standard projective geometry:

$20$4

and the benchmark also provides planar homography support for road-plane mappings:

$20$5

The included calibration assets are explicitly described as compatible with the cameratransform package and support projection from image coordinates to geographic coordinates (Lin et al., 11 Jul 2025).

RoundaboutHD’s task protocols follow established evaluation conventions. Object detection uses [email protected], [email protected], [email protected], and mAP. Single-camera tracking reports IDF1, IDP, IDR, MOTA, and ID switches. Image-based ReID reports CMC Rank-$20$6, mAP, and mINP. Multi-camera tracking aligns with the CVPR AI City Challenge and emphasizes global association quality through IDF1 with IDP and IDR (Lin et al., 11 Jul 2025).

4. HD maps, adjacent corpora, and generated roundabout benchmarks

The broader RoundaboutHD ecosystem includes several resources that are not identically named, but that instantiate the same conjunction of roundabout geometry, trajectories, and HD map structure.

openDD is the clearest large-scale example. It provides trajectories for all road users present in a scene together with HD map data for seven different roundabouts. Data were recorded from a DJI Phantom 4 at $20$7 and $20$8, stabilized and rectified, and released as $20$9 accurately tracked trajectories collected over $H$0 drone flights and $H$1 hours. Each site includes lane centerlines, lane boundaries, drivable areas, shapefiles $H$2, an XML mirror of the shapefile content, and a geo-referenced and anonymized reference image. Trajectories are given in UTM coordinates and include position, yaw, width, length, speed, lateral acceleration, tangential acceleration, total acceleration, and class labels $H$3 (Breuer et al., 2020).

Procedural map generation extends the concept into simulation. The roundabout-generation paper produces classic and turbo roundabouts directly in OpenDRIVE from incident-road tuples containing position, heading, and per-side lane counts. It fits a least-squares circle to the incident points and selects the initial ring radius as

$H$4

A center-offset term, defined as the angular difference between the incident road’s heading and the radial direction from roundabout center to the incident point, drives ring-segment selection and straight-approach length. The ring can then be distorted through Perlin-noise perturbation and connected by parametric cubic roads, allowing non-perfectly circular roundabouts that remain OpenDRIVE-compliant (Ikram et al., 2023).

A separate CARLA-based line of work contributes an HD roundabout benchmark for articulated vehicles rather than for perception or tracking. That work defines five custom roundabouts in RoadRunner with diameters $H$5, $H$6, $H$7, $H$8, and $H$9, uses HD lane centerline waypoints as route definitions, trains on $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$0 routes from the $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$1, $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$2, and $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$3 roundabouts, and tests on $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$4 routes from the unseen $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$5 and $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$6 roundabouts. With a twin-Q soft actor-critic policy, the reported overall test success rate is $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$7 (Attard et al., 2024).

Taken together, these resources show that “RoundaboutHD” is not solely a tracking benchmark label. It also denotes, in a broader technical sense, HD-map–enabled roundabout environments that support localization, behavior analysis, prediction, or control.

5. Evaluation regimes, baselines, and real-time systems

The 2025 RoundaboutHD benchmark was accompanied by baseline results across detection, single-camera tracking, image-based ReID, and multi-camera tracking. For detection, using inference settings of confidence $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$8, NMS enabled, image size $$(x, y) = \left(M^{(1,..,P)}_{lon}(i,j), M^{(1,..,P)}_{lat}(i,j)\right).$$9, and IoU threshold $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$0, YOLOv11x achieves mAP values of $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$1, $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$2, $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$3, and $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$4 on cam01–cam04, while YOLOv12x achieves $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$5, $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$6, $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$7, and $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$8. The paper notes that YOLOv11x slightly outperforms YOLOv12x $$r(\theta) = k_1 \theta + k_2 \theta^3 + k_3 \theta^5 + \ldots$$9 average mAP), and that static parked vehicles visible in some scenes, especially cam01, are excluded in ground truth but detected by YOLO, inflating false positives (Lin et al., 11 Jul 2025).

Single-camera tracking baselines are reported in the BoxMOT framework using YOLOv12x detections and stationary-tracklet filtering. Averaged across four cameras, ByteTrack achieves IDF1 $K$0 and MOTA $K$1, BotSort achieves IDF1 $K$2 and MOTA $K$3, DeepOCSort achieves IDF1 $K$4 and MOTA $K$5, OCSort achieves IDF1 $K$6 and MOTA $K$7, and BoostTrack achieves IDF1 $K$8 and MOTA $K$9. The paper highlights ByteTrack as the best average performer on IDF1 and MOTA. In image-based ReID, FastReID models pretrained only on ImageNet perform poorly, but after fine-tuning on the RoundaboutHD ReID subset, SBS reaches mAP $360^\circ$00, Rank-1 $360^\circ$01, Rank-5 $360^\circ$02, and mINP $360^\circ$03. By contrast, multi-camera tracking remains difficult: the ELECTRICITY baseline yields IDF1 $360^\circ$04, IDP $360^\circ$05, and IDR $360^\circ$06, markedly below CityFlow and Synthehicle (Lin et al., 11 Jul 2025).

A later systems paper uses RoundaboutHD as the primary evaluation dataset for a scalable, edge-centric MCVT pipeline. SAE-MCVT deploys object detection, single-camera tracking, geo-mapping, and feature extraction on edge devices, transmits only lightweight metadata to a central workstation, and performs cross-camera association using a self-supervised camera link model. On RoundaboutHD, the edge-side detector is YOLOv11n retrained on the detection subset with confidence threshold $360^\circ$07, NMS IoU $360^\circ$08, and inference size $360^\circ$09; the tracker is ByteTrack; geo-mapping uses Cameratransform with an assumed vehicle height of $360^\circ$10; and the feature extractor is a ResNet-50 pretrained on the image-based ReID subset. The association cost is written as

$360^\circ$11

with RoundaboutHD settings $360^\circ$12 and $360^\circ$13. The reported identification metrics are IDF1 $360^\circ$14, IDP $360^\circ$15, and IDR $360^\circ$16. The latency table reports $360^\circ$17 per frame for object detection, $360^\circ$18 for single-camera tracking, and $360^\circ$19 for geo-mapping, while feature extraction and cross-camera association show no queue overflow (Lin et al., 17 Nov 2025).

These results establish a characteristic asymmetry of the benchmark: per-camera perception and ReID can be very strong after domain adaptation, but global association across non-overlapping views remains substantially harder.

6. Limitations, misconceptions, and future directions

The first limitation is terminological. It is incorrect to assume that RoundaboutHD always denotes the 2025 benchmark. In the Ann Arbor perception paper, the term is absent from the paper text and must be understood functionally through the deployed HD-map–anchored system and its data stream; in openDD, it is likewise a descriptive interpretation rather than a dataset name (Zou et al., 2022, Breuer et al., 2020).

The second limitation is dataset scope. The 2025 benchmark is a single geographic scene with moderate duration, and weather variations and nighttime conditions are not explicitly cataloged. The Ann Arbor deployment is spatially richer in the sense of continuous $360^\circ$20 operation, but its calibration is manual, depends on Google Maps landmark annotation, assumes stationary cameras and a planar road surface, and exhibits higher out-ROI errors. The paper also notes performance degradation with camera shake in one windy trip, the absence of explicit appearance re-identification, and the lack of a public release of RoundaboutHD/AA-Fisheye/AA-Thermal data, links, splits, or licensing terms (Lin et al., 11 Jul 2025, Zou et al., 2022).

The third limitation concerns what HD roundabout resources usually omit. openDD does not enumerate exact geographic locations, roundabout size, number of arms, or single-versus-multi-lane status for each site. The procedural-generation work does not explicitly enforce formal design-code constraints such as minimum radius by speed class, maximum curvature, minimum weaving length, or explicit conflict-point analysis. The tractor–trailer CARLA benchmark provides no public URL for maps or scenarios, and its experiments do not include multi-agent traffic interactions (Breuer et al., 2020, Ikram et al., 2023, Attard et al., 2024).

Future directions are correspondingly specific. The Ann Arbor deployment identifies automated calibration, explicit cross-camera synchronization, improved fusion with uncertainty-aware weighting, and broader public dataset release as extensions. The 2025 RoundaboutHD benchmark identifies broader lighting, weather, and traffic patterns, and potentially more camera placements to widen scenario diversity. A plausible implication is that future “RoundaboutHD” resources will be judged less by the presence of HD geometry alone than by how tightly geometry, calibration, cross-camera identity, and long-horizon temporal coverage are integrated into a single reproducible benchmark (Zou et al., 2022, Lin et al., 11 Jul 2025).