Pith. sign in

REVIEW 5 major objections 5 minor 31 references

CLIFE: Camera-LiDAR Fusion Framework for Edge-Deployable Roadside VRU Perception

T0 review · 5 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a camera-LiDAR fusion pipeline running entirely on one embedded device can keep continuous real-time tracks of pedestrians, cyclists, scooterists, and wheelchair users at urban intersections, without cloud offloading.

desk verdict A serious edge-deployment effort whose headline robustness claim collides with its own light-rain results. read the letter →

arxiv 2607.16154 v1 pith:UW7HOEBQ submitted 2026-07-17 cs.CV cs.SYeess.SY

Tam Bang , Hoang H. Nguyen , Lei Cheng , Lihao Guo , Siyang Cao , Hussam Abubakr , Tianya Zhang , Austin Harris
show 1 more author
Mina Sartipi
This is my paper
classification cs.CVcs.SYeess.SY
keywords camera-LiDARfusionroadsideperceptionvulnerableroadusersedgecomputinglatetargetlesscalibrationmulti-objecttrackingintersectionsafety
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that reliable roadside perception of vulnerable road users can be achieved entirely on an edge device by fusing camera and LiDAR streams, eliminating the need for cloud offloading. It presents CLIFE, a pipeline that couples targetless online calibration with a lightweight late-fusion tracker. The paper reports that fusion improves tracking accuracy and identity consistency over either sensor alone in sunny and cloudy conditions, and that LiDAR maintains tracks when the camera is occluded or dazzled. The practical significance is that, if correct, traffic agencies could deploy continuous multi-object VRU tracking at intersections without server-grade infrastructure or manual recalibration.

What carries the argument

The load-bearing object is a 3x3 ground-plane homography H that maps LiDAR ground coordinates to image coordinates. It is estimated by a targetless calibration module that matches cross-modal detections using learned spatial positions, appearance embeddings, and semantic labels, then refines alignment by minimizing reprojection error. Fusion projects LiDAR 3D boxes through H, queries a KD-tree for camera detections within a pixel radius, and greedily assigns one-to-one matches per object class (VRUs and vehicles). A FIFO tracker accumulates fused, camera-only, and LiDAR-only detections over K frames, using miss counters to survive single-sensor dropouts. The combination of KD-tree search and

What would settle it

Run the CLIFE pipeline, with calibration fine-tuned at one intersection, at a second deployed intersection without any per-site fine-tuning; if MOTA/IDF1 drops markedly or calibration AED exceeds a few pixels, the scalability and generalization claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that a ground-plane homography and a distance-based late-fusion stage, both running on one embedded device, can produce continuous and accurate VRU tracks at a real intersection. Calibration is performed on demand: a targetless module matches cross-modal detections using spatial, appearance, and semantic cues, then refines the homography to a reported 16.02-pixel average Euclidean distance. Fusion projects LiDAR detections into the image plane, associates them with camera detections per category via KD-tree radius search and greedy matching, and runs in O(N log N) time per frame, sustaining 53.2 FPS for the fusion core on an embedded GPU. In quantitative evaluati

Load-bearing premise

The claim of 12-intersection scalability rests on the assumption that the per-site fine-tuned calibration and camera models transfer to new intersections without re-tuning; the paper quantitatively evaluates only a single intersection.

Editorial extensions

If this is right

  • A single embedded node can serve two full camera-LiDAR intersections at roughly 10 FPS per stream, so multi-intersection corridors can be scaled by adding edge nodes in parallel.
  • Fusion keeps a single consistent track through occlusion, glare, and heavy rain where either sensor alone loses the object, lowering the risk of identity fragmentation in downstream safety applications.
  • On-demand targetless recalibration removes the need for periodic manual calibration, allowing long-term unattended deployment after an initial ~30-minute on-site refinement.
  • Because the fusion stage operates only on structured detection outputs once alignment is established, the core tracking does not require raw sensor data, reducing bandwidth and memory pressure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported light-rain results, where LiDAR-only outperforms fusion, suggest that a condition-aware sensor selector could further improve accuracy; the authors leave this as future work, but their own numbers imply the gain.
  • The calibration module was fine-tuned to the specific intersection geometry; a controlled test across the 12 deployed sites would clarify whether the claimed scalability holds without per-site retraining.
  • The pipeline's dependence on a proprietary LiDAR perception API limits end-to-end optimization; the on-device PointPillars variant hints that a fully open, reproducible edge deployment is within reach.
  • The quantitative evaluation at a single intersection means the '12 intersections' claim is a deployment statement, not yet a generalization statement; measuring MOTA/IDF1 at a second site would settle it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. CLIFE is an edge-native camera–LiDAR late-fusion framework for roadside VRU perception. Phase 1 uses a targetless homography-based calibration module (CalibRefine) to align LiDAR ground-plane projections with the camera image; Phase 2 projects LiDAR detections into image coordinates, associates them with camera detections via KD-tree radius search and greedy one-to-one assignment, and maintains fused, camera-only, and LiDAR-only hypotheses in a FIFO-based tracker. The paper reports deployment on 12 signalized intersections in Chattanooga, detailed quantitative evaluation at one intersection (Georgia Ave. and M.L.K. Blvd.) on 9,000 synchronized frame pairs spanning sunny, cloudy, and light-rain conditions, and qualitative results for heavy rain and nighttime. The headline claims are that fusion 'substantially enhances the perceptual range and robustness' over individual sensors, that the pipeline runs entirely on an NVIDIA Jetson AGX Thor with the late-fusion core at 53.2 FPS, and that one edge node supports five parallel CLIFE workers.

Significance. If fully substantiated, the paper would be a useful practical contribution: an edge-deployable fusion pipeline with targetless recalibration, bounded O(N log N) per-frame cost, real-world intersection data, and a direct three-way comparison of camera-only, LiDAR-only, and fused tracking. The deployment on real intersections and the throughput/scalability measurements are concrete strengths, as is the modular architecture that supports independent component upgrades. However, the central robustness claim is only partially supported by the reported numbers: under light rain, the fused tracker is worse than LiDAR-only, and the 'perceptual range' claim is not quantitatively evaluated at all. The evaluation also depends on a site-tuned calibration module from the same group with no external calibration baseline. These are fixable with revised claims, additional analyses, and broader evaluation, so the contribution is salvageable; as written, the evidence does not fully support the abstract and introduction.

major comments (5)
  1. [§IV-D3, Table V] The light-rain row directly contradicts the paper's central claim. Under light rain, fusion achieves MOTA 65.8 and IDF1 71.2, while LiDAR-only achieves 70.1 and 78.3 — a 4.3-point MOTA and 7.1-point IDF1 deficit. Fusion is therefore the worst of the three configurations in this condition, not a 'substantial enhancement' in robustness. The sentence in §IV-D3 that 'condition-aware sensor selection may improve tracking performance' concedes that the fixed fusion architecture is not robust under a condition explicitly included in the evaluation. Since the abstract and §I claim robustness 'under varied environmental and traffic conditions,' this is a load-bearing inconsistency. Please either modify the fusion to handle rain-induced degradation and re-evaluate, or revise the claims and add a condition-selection mechanism; in either case, report per-sequence variability so the reader can judge
  2. [§IV-D2, §IV-A, §V] The calibration evaluation is circular with respect to the fusion evaluation. CalibRefine (ref [26]) is a prior paper by overlapping authors, and the module was fine-tuned on 'our intersection-specific calibration subset' at the same intersection used for end-to-end evaluation. The only reported calibration accuracy is AED 16.02 px, with no external baseline (e.g., manual/checkerboard calibration, an alternative targetless method, or an error analysis). Section V itself states that performance 'remains sensitive to extrinsic calibration accuracy.' Consequently, the fusion gains in Table V cannot be cleanly attributed to the late-fusion algorithm rather than to site-tuned calibration. Please add an external calibration comparison and a sensitivity or transfer analysis across at least one additional intersection.
  3. [Abstract; §I; §IV-C] The claimed enhancement of 'perceptual range' is not measured. Section IV-C defines metrics for detection (precision/recall/mAP), tracking (MOTA/IDF1), and calibration (AED), but no distance- or range-dependent metric appears anywhere in the quantitative evaluation. Figure 7 and §IV-D5 provide only qualitative occlusion/range examples. To support the range claim, report a distance-based evaluation (e.g., tracking recall or MOTA in range bins from the sensor) for camera-only, LiDAR-only, and fusion.
  4. [§IV-D3, §IV-B] The quantitative evidence lacks measures of variability or replication. The 9,000 frame pairs are described as 3,000 per weather condition, but the text does not report the number of independent episodes, object trajectories, or sequences within each condition. Several headline differences are small (e.g., sunny MOTA 78.6 vs. 76.4; cloudy MOTA 65.4 vs. 63.2), and without per-sequence metrics, standard deviations, or confidence intervals, it is not possible to assess whether these differences are meaningful. This is particularly important for the light-rain result, where the reported deficit is also unaccompanied by any uncertainty estimate. Please provide per-sequence statistics and object-level counts.
  5. [§IV-A, §IV-Ba, §I] The 12-intersection deployment claim is not supported by the quantitative evaluation. Only one intersection is evaluated in depth, the camera perception model is fine-tuned on frames from two intersections, and CalibRefine is fine-tuned on a site-specific subset. No cross-site evaluation or ablation shows whether per-site retuning is required. As a result, 'deployed across 12 signalized intersections' is an operational assertion rather than a demonstrated generalization property. Please provide at least one cross-intersection tracking comparison, or explicitly state the per-site setup effort and its limits.
minor comments (5)
  1. [Algorithm 1, line 14] The Fuse(c, ℓ) function is not formally defined. In particular, the paper says 'ties favor the camera' for class/confidence, but does not specify how class conflicts are resolved when the camera and LiDAR labels disagree (e.g., camera 'wheelchair user' vs. LiDAR 'pedestrian'). Please define the attribute-fusion rules precisely.
  2. [Table VI, §IV-D4] The 'LiDAR (on-device)' throughput numbers exclude 'the additional ~97 ms delay for transferring LiDAR point clouds from the Ouster device.' This caveat should appear in the table caption, since the on-device row is otherwise easily read as a direct point-cloud-to-tracking latency figure.
  3. [§IV-B] The dataset description reports 3,000 frame pairs per weather condition, but it is not clear whether these are temporally contiguous clips or scattered frames. Please clarify the temporal structure and how the train/evaluation split avoids frame-level leakage, especially since the camera model is trained on data from the same intersections.
  4. [§II, References] Reference [19] is a product web page; for the LiDAR specifications cited in Table II (e.g., 'range up to ~200 m'), a datasheet citation would be more appropriate and verifiable.
  5. [§IV-D5] The qualitative results under heavy rain and nighttime are useful, but the paper should state explicitly that these conditions are not part of the quantitative comparison, so that readers do not over-interpret Fig. 7 as quantitative evidence for robustness.

Circularity Check

2 steps flagged · score 4.0 of 10

Calibration accuracy is the optimization loss itself, and the calibration contribution rests on a self-cited module fine-tuned on the evaluation site; the fusion comparison itself is independent.

  1. self definitional [Section III-B (Phase 1), Section IV-C (Evaluation Metrics), Section IV-D2 (Calibration)]
    "Given a batch of time-synchronized camera–LiDAR pairs S, the module estimates a ground-plane homography H ... followed by iterative geometric refinement to minimize reprojection error. ... calibration accuracy is quantified by the Average Euclidean Distance (AED) in pixels, measured between H-projected LiDAR ground points and their matched camera object center, averaged across all synchronized pairs."

    CalibRefine solves for H by minimizing the reprojection error between H-projected LiDAR points and matched camera object centers; AED is then defined as exactly that reprojection error (average Euclidean distance between H-projected LiDAR ground points and matched camera object centers). The reported 16.02 px AED is therefore the optimization loss, not an independent measurement of extrinsic accuracy, since no surveyed ground-truth extrinsic is used. The calibration-accuracy claim is self-consistent by construction.

  2. self citation load bearing [Section IV-D2 (Calibration)]
    "We use CalibRefine [26], a targetless, on-device calibration module that estimates the camera–LiDAR homography by matching cross-modal detections through learned spatial, appearance, and semantic cues, followed by iterative geometric refinement. The module’s cross-modal matching network was initialized from a pre-trained checkpoint and fine-tuned on our intersection-specific calibration subset to adapt to the roadside sensor geometry and mounting configuration, achieving 94.98% cross-modal pairing accuracy."

    Reference [26] is prior work by the same authors (Cheng, Guo, Zhang, Bang, Harris, Sartipi, Cao) and is the stated basis for the calibration module, a key contribution of CLIFE. The module is fine-tuned on the same intersection used for the end-to-end fusion evaluation, and its only quantitative validation is the self-referential AED metric above. The paper provides no independent or external calibration baseline, so this load-bearing calibration component rests on an unverified self-citation plus a check that reduces to its own training objective.

full rationale

CLIFE's late-fusion method (Algorithm 1) is otherwise an independent baseline comparison: camera-only, LiDAR-only, and fusion use the same detection streams, and the fusion equations do not encode the outcome. The light-rain deficit in Table V and the lack of cross-site generalization tests are correctness/generalization concerns, not circularity. The circular component is confined to calibration. CalibRefine [26], by overlapping authors, is adopted as Phase 1 and fine-tuned on the same Georgia & M.L.K. intersection used for evaluation. The claimed calibration accuracy (AED 16.02 px) is measured as the distance between H-projected LiDAR points and matched camera centers, which is exactly the reprojection error minimized when estimating H; no surveyed ground-truth extrinsic is used. Thus the calibration result is self-consistent by construction, and the self-cited prior work carries the calibration contribution without independent external validation. Since the central fusion-robustness claim is otherwise supported by within-site comparisons of independent sensor streams, the circularity is partial rather than total.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The system is built from off-the-shelf perception components (YOLOv11, ByteTrack, PointPillars, AB3DMOT) and the authors' prior calibration work; the only hand-set parameters are the fusion/tracking constants and the fine-tuning of CalibRefine to the test intersection. The main assumptions are the planarity of the ground plane, the reliability of the proprietary BlueCity API, and the representativeness of single-intersection evaluation for the 12-intersection deployment claim.

free parameters (5)
  • Search radius r_s = 100 px
    KD-tree radius search threshold for matching camera and LiDAR detections; hand-chosen, not ablated.
  • Miss threshold tau_miss = 10 frames
    Number of consecutive misses before track termination; hand-set.
  • N_consec = 5
    Consecutive frames for both-sensor-miss track termination; hand-set.
  • FIFO buffer length K = 30
    Number of recent frames used for track association; hand-set.
  • CalibRefine fine-tuning = Intersection-specific calibration subset
    The cross-modal matching network is fine-tuned on the same intersection later used for evaluation; this is fitting to the test environment.
assumptions (5)
  • domain assumption The ground plane is locally planar for homography estimation
    Phase 1 estimates a homography mapping road-plane coordinates to image coordinates; if the road is not planar, projection errors increase.
  • domain assumption BlueCity LiDAR API provides correct 10 Hz detections at all 12 intersections
    The end-to-end system relies on the proprietary Ouster BlueCity backend for LiDAR perception; the paper does not validate this component independently.
  • domain assumption Camera–LiDAR synchronization at 10 Hz is sufficient for fusion
    Camera frames are downsampled to match LiDAR's 10 Hz; higher-frequency camera data is discarded, and the impact is not analyzed.
  • ad hoc to paper Single-intersection evaluation generalizes to 12-intersection deployment
    Quantitative evaluation is conducted at one representative intersection; the 12-intersection deployment claim assumes no site-specific re-tuning is needed.
  • domain assumption Manual annotation is accurate enough for metric computation
    Annotations were auto-labeled then manually refined; no inter-annotator agreement or quality checks are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLIFE: Camera-LiDAR Fusion Framework for Edge-Deployable Roadside VRU Perception." pith.science (2026). https://pith.science/paper/UW7HOEBQ

@misc{pith2026260716154,
  author       = {Pith},
  title        = {Pith review of: CLIFE: Camera-LiDAR Fusion Framework for Edge-Deployable Roadside VRU Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UW7HOEBQ}},
  note         = {Machine review of arXiv:2607.16154}
}
read the original abstract

Reliable roadside perception of vulnerable road users (VRUs) remains challenging under occlusions, variable lighting, and diverse weather conditions, particularly under strict edge-computing and latency constraints. Existing multi-sensor fusion systems rely on cloud or server-grade infrastructure, creating a deployment gap at real-world intersections. We present CLIFE, an edge-native camera-LiDAR fusion framework that integrates targetless online calibration and lightweight late-fusion tracking entirely on a single embedded device, without cloud offloading. CLIFE adaptively refines camera-LiDAR alignment on demand and performs multi-sensor fusion and track association with O(N log N) per-frame cost. We deploy CLIFE across 12 signalized intersections in Chattanooga and conduct an in-depth evaluation at a representative intersection using synchronized camera-LiDAR data that spans diverse daytime, nighttime, and weather conditions. Our experiments demonstrate that the fusion architecture substantially enhances the perceptual range and robustness of the individual sensors under varied environmental and traffic conditions. The late-fusion core operates at 53.2 FPS on the Jetson AGX Thor, ensuring high throughput for real-time intersection-scale applications. By centering perception at the edge, CLIFE provides a deployable foundation for downstream safety applications, while reducing bandwidth and calibration overhead for agencies operating multi-intersection corridors.

Figures

Figures reproduced from arXiv: 2607.16154 by the authors.

Figure 1
Figure 1. Roadside VRU perception challenges: (a) camera [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CLIFE architecture. The framework [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of homography transformation. With accu [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Street-view of the sensor setup. Each intersection is equipped with an Ouster BlueCity backend [19], a co-located edge system that runs on-site, time￾aligns and merges point clouds from the northwest (NW) and southeast (SE) LiDAR units, and delivers structured percepti…
Figure 5
Figure 5. Figure 5: Sample images from the multi-sensor VRU dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: LiDAR ground-plane projections overlaid on the orig [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results demonstrating the robustness of our fusion method under challenging real-world scenarios. LiDAR [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 6 linked inside Pith

  1. [26]

    Calibrefine: Deep learning-based online automatic target- less lidar-camera calibration with iterative and attention-driven post- refinement,

    L. Cheng, L. Guo, T. Zhang, T. Bang, A. Harris, M. Hajij, M. Sartipi, and S. Cao, “Calibrefine: Deep learning-based online automatic target- less lidar-camera calibration with iterative and attention-driven post- refinement,”arXiv preprint arXiv:2502.17648, 2025

    arXiv 2025

  2. [1]

    Vulnerable road user research plan,

    Federal Highway Administration, “Vulnerable road user research plan,” U.S. Department of Transportation, Federal Highway Administration, Washington, DC, Tech. Rep., July 2023. [Online]. Available: https://highways.dot.gov/sites/fhwa.dot.gov/files/ FHW A-Vulnerable-Road-User.pdf

    2023

  3. [2]

    Fatality facts 2023: State by state,

    Insurance Institute for Highway Safety, “Fatality facts 2023: State by state,” https://www.iihs.org/topics/fatality-statistics/detail/state-by-state, July 2023

    2023

  4. [3]

    Vulnerable road user detection and safety enhancement: A comprehensive survey,

    R. M. Silva, G. F. Azevedo, M. V . Berto, J. R. Rocha, E. C. Fidelis, M. V . Nogueira, P. H. Lisboa, and T. A. Almeida, “Vulnerable road user detection and safety enhancement: A comprehensive survey,”Expert Systems with Applications, p. 128529, 2025

    2025

  5. [4]

    Multi-view 3d object detection network for autonomous driving,

    X. Chen, H. Ma, J. Wan, B. Li, and T. Xia, “Multi-view 3d object detection network for autonomous driving,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1907–1915

    2017

  6. [5]

    Multimodal virtual point 3d detection,

    T. Yin, X. Zhou, and P. Kr ¨ahenb¨uhl, “Multimodal virtual point 3d detection,” inAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  7. [6]

    Infradet3d: Multi-modal 3d object detection based on roadside infrastructure camera and lidar sensors,

    W. Zimmer, J. Birkner, M. Brucker, H. T. Nguyen, S. Petrovski, B. Wang, and A. C. Knoll, “Infradet3d: Multi-modal 3d object detection based on roadside infrastructure camera and lidar sensors,” in2023 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2023, pp. 1–8

    2023

  8. [7]

    Rp3d: A roadside perception framework for 3d object detection via multi-view sensor fusion,

    S. Zheng, R. Huang, Y . Ji, M. Ye, and W. Li, “Rp3d: A roadside perception framework for 3d object detection via multi-view sensor fusion,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 16–33

    2024

Show all 31 references
  1. [8]

    Vips: Real-time perception fusion for infrastructure-assisted autonomous driv- ing,

    S. Shi, J. Cui, Z. Jiang, Z. Yan, G. Xing, J. Niu, and Z. Ouyang, “Vips: Real-time perception fusion for infrastructure-assisted autonomous driv- ing,” inProceedings of the 28th annual international conference on mobile computing and networking, 2022, pp. 133–146

    2022

  2. [9]

    Clocs: Camera-lidar object candi- dates fusion for 3d object detection,

    S. Pang, D. Morris, and H. Radha, “Clocs: Camera-lidar object candi- dates fusion for 3d object detection,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 10 386–10 393

    2020

  3. [10]

    Multi-task multi- sensor fusion for 3d object detection,

    M. Liang, B. Yang, Y . Chen, R. Hu, and R. Urtasun, “Multi-task multi- sensor fusion for 3d object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 7345– 7353

    2019

  4. [11]

    Multi-sensor fusion for multi- target detection and tracking,

    L. Liu, J. Tian, Z. Shi, J. Fan, and Y . Rui, “Multi-sensor fusion for multi- target detection and tracking,” inAutonomous Vehicles and Systems. River Publishers, 2024, pp. 175–217

    2024

  5. [12]

    Survey of extrinsic calibration on lidar-camera system for intelligent vehicle: Challenges, approaches, and trends,

    P. An, J. Ding, S. Quan, J. Yang, Y . Yang, Q. Liu, and J. Ma, “Survey of extrinsic calibration on lidar-camera system for intelligent vehicle: Challenges, approaches, and trends,”IEEE Transactions on Intelligent Transportation Systems, 2024

    2024

  6. [13]

    Automatic targetless lidar–camera calibration: a survey,

    X. Li, Y . Xiao, B. Wang, H. Ren, Y . Zhang, and J. Ji, “Automatic targetless lidar–camera calibration: a survey,”Artificial Intelligence Review, vol. 56, no. 9, pp. 9949–9987, 2023

    2023

  7. [14]

    Real time object detection using lidar and camera fusion for autonomous driving,

    H. Liu, C. Wu, and H. Wang, “Real time object detection using lidar and camera fusion for autonomous driving,”Scientific Reports, vol. 13, no. 1, p. 8056, 2023

    2023

  8. [15]

    Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,”arXiv preprint arXiv:2205.13542, 2022

    2022 arXiv

  9. [16]

    Deep structural information fusion for 3d object detection on lidar–camera system,

    P. An, J. Liang, K. Yu, B. Fang, and J. Ma, “Deep structural information fusion for 3d object detection on lidar–camera system,”Computer Vision and Image Understanding, vol. 214, p. 103295, 2022

    2022

  10. [17]

    Fast-clocs: Fast camera-lidar object candidates fusion for 3d object detection,

    S. Pang, D. Morris, and H. Radha, “Fast-clocs: Fast camera-lidar object candidates fusion for 3d object detection,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 187–196

    2022

  11. [18]

    Infrastructure-assisted 3d detection networks based on camera-lidar early fusion strategy,

    J. Yao, J. Zhou, Y . Wang, Z. Gao, and W. Hu, “Infrastructure-assisted 3d detection networks based on camera-lidar early fusion strategy,” Neurocomputing, vol. 600, p. 128180, 2024

    2024

  12. [19]

    Ouster BlueCity: AI-Driven Lidar for Smarter Cities,

    Ouster Inc., “Ouster BlueCity: AI-Driven Lidar for Smarter Cities,” https://ouster.com/products/software/bluecity, 2025, accessed: Nov. 1, 2025

    2025

  13. [20]

    Computer Vision Annotation Tool (CV AT),

    CV AT.ai Corporation, “Computer Vision Annotation Tool (CV AT),” Nov. 2023. [Online]. Available: https://github.com/cvat-ai/cvat

    2023

  14. [21]

    Sustech points: A portable 3d point cloud interactive annotation platform system,

    E. Li, S. Wang, C. Li, D. Li, X. Wu, and Q. Hao, “Sustech points: A portable 3d point cloud interactive annotation platform system,” in2020 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2020, pp. 1108–1115

    2020

  15. [22]

    NVIDIA TensorRT: High-Performance Deep Learning Inference SDK,

    NVIDIA Corporation, “NVIDIA TensorRT: High-Performance Deep Learning Inference SDK,” https://github.com/NVIDIA/TensorRT, 2025

    2025

  16. [23]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  17. [24]

    Bytetrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, “Bytetrack: Multi-object tracking by associating every detection box,” inEuropean conference on computer vision. Springer, 2022, pp. 1–21

    2022

  18. [25]

    Bot-sort: Robust associa- tions multi-pedestrian tracking,

    N. Aharon, R. Orfaig, and B.-Z. Bobrovsky, “Bot-sort: Robust associa- tions multi-pedestrian tracking,”arXiv preprint arXiv:2206.14651, 2022

    2022 arXiv

  19. [27]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

    2019

  20. [28]

    Lidar AI Solution,

    NVIDIA AI IOT, “Lidar AI Solution,” https://github.com/ NVIDIA-AI-IOT/Lidar AI Solution, 2023, accessed: 2026

    2023

  21. [29]

    Ab3dmot: A baseline for 3d multi-object tracking and new evaluation metrics,

    X. Weng, J. Wang, D. Held, and K. Kitani, “Ab3dmot: A baseline for 3d multi-object tracking and new evaluation metrics,”arXiv preprint arXiv:2008.08063, 2020

    2008 arXiv

  22. [30]

    Urbaning- v2x: A large-scale multi-vehicle, multi-infrastructure dataset across multiple intersections for cooperative perception,

    K. C. Sekaran, M. Geisler, D. R ¨oßle, A. Mohan, D. Cremers, W. Utschick, M. Botsch, W. Huber, and T. Sch ¨on, “Urbaning- v2x: A large-scale multi-vehicle, multi-infrastructure dataset across multiple intersections for cooperative perception,”arXiv preprint arXiv:2510.23478, 2025

    2025

  23. [31]

    R-livit: A lidar-visual-thermal dataset enabling vulnerable road user focused roadside perception,

    J. Mirlach, L. Wan, A. Wiedholz, H. E. Keen, and A. Eich, “R-livit: A lidar-visual-thermal dataset enabling vulnerable road user focused roadside perception,”arXiv preprint arXiv:2503.17122, 2025

    2025 arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.