Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Label-Free Long-Horizon 3D UAV Trajectory Prediction via Motion-Aligned RGB and Event Cues

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · 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 label-free, video-only pipeline can forecast a drone's 3D position five seconds ahead with roughly 40 percent lower error than supervised image-only and audio-visual baselines.

desk verdict The label-free pipeline's real gain over the best supervised baseline is ~1.4%, not the 40% claimed in the abstract — worth a revision, not a desk reject. read the letter →

arxiv 2507.03365 v1 pith:C2LPOD77 submitted 2025-07-04 cs.RO

Hanfang Liang , Shenghai Yuan , Fen Liu , Yizhuo Yang , Bing Wang , Zhuyu Huang , Chenyang Shi , Jing Jin This is my paper · ORCID
classification cs.RO
keywords UAVtrajectorypredictionself-supervisedlearningevent-basedvisionLiDARpseudo-labelingVision-Mambacross-modalalignmentcounter-dronesystemslong-horizonforecasting
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 3D UAV trajectories can be forecast five seconds ahead without any manual 3D labels, using video plus automatically generated pseudo-labels. It builds those pseudo-labels by extracting smooth motion trajectories from raw LiDAR and radar point clouds with a Temporal-KNN gradient filter, then aligning them with RGB and simulated-event motion cues in the fisheye image plane. A stage-wise self-supervised Vision-Mamba network is trained on the aligned pseudo-labels to detect the drone and regress its 3D position and future trajectory. On the MMAUD V1/V2 dataset the paper reports a 5-second RMSE of 6.20 m, roughly 40 percent lower than supervised image-only and audio-visual baselines, with no LiDAR or radar needed at inference. The motivation is practical: a 5-second lead time is enough to activate countermeasures such as lasers, nets, or interceptor drones, so accurate label-free forecasting would make anti-drone systems more proactive.

What carries the argument

The central mechanism is the Temporal-KNN gradient filter: it links each LiDAR/radar point to its K nearest neighbors in subsequent frames, records motion vectors, computes a mean gradient magnitude per point, and keeps only vectors with gradient below a smoothness threshold tau, so the drone's coherent motion survives while static clutter is rejected. The second load-bearing piece is cross-modal motion alignment, which projects the surviving 3D trajectory vectors into fisheye image coordinates and matches them against ORB keypoint tracks and simulated event cues using a cosine-similarity plus full-state alignment loss on position, velocity, and acceleration. The third piece is the stage-wise self-supervised architecture: a Vision-Mamba encoder with dual RGB/event branches and cross-attention fusion, a DETR-style decoder for detection, and two feedforward heads that first map image-space detections to 2D motion states and then to 3D states, with pseudo-labels supervising each stage.

What would settle it

Run the Temporal-KNN pseudo-label generator on a sequence where the drone flies near the ground or beyond 100 meters, compare the extracted labels against high-accuracy survey ground truth, and test whether varying the smoothness threshold tau over a reasonable range changes the 5-second prediction RMSE; if a single fixed tau cannot keep the drone while rejecting clutter in those regimes, the label-free claim fails for exactly the cases the paper's own failure analysis names.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that 3D UAV trajectory prediction for counter-drone lead time can be done without any manual 3D labels: an unsupervised Temporal-KNN gradient filter extracts smooth motion trajectories from asynchronous raw LiDAR and radar point clouds, cross-modal alignment with ORB and simulated event motion in the fisheye image plane turns them into pseudo-labels, and a stage-wise self-supervised Vision-Mamba network trained on RGB plus event cues predicts future 3D positions. On the MMAUD V1/V2 sequences the reported result is 6.20 m RMSE at the 5-second horizon, about 40 percent lower than supervised image-only and audio-visual baselines, with no LiDAR or radar needed at inference. The paper also states that simulated event cues are a conservative lower bound, so real event sensors should improve rather than degrade performance.

Load-bearing premise

The entire pipeline depends on the fixed smoothness threshold tau in the Temporal-KNN filter correctly separating the drone's LiDAR and radar points from background noise; when the drone is occluded, near large objects, or very far away, that separation fails and every pseudo-label inherits the error.

Editorial extensions

If this is right

  • Counter-drone systems could act on a 5-second future 3D position using only a monocular video feed, removing the need for LiDAR or radar at deployment.
  • New deployment sites would not require manual 3D annotation: the same Temporal-KNN and motion-alignment pipeline can generate pseudo-labels from raw LiDAR and RGB data alone.
  • Because simulated event cues are treated as a conservative lower bound, switching to real wide-field-of-view event cameras should improve motion fidelity and prediction accuracy without changing the architecture.
  • The stage-wise self-supervised design separates detection, 2D-to-3D projection, and motion forecasting, so failures in early stages can be isolated and each stage can be improved independently.

Reading between the lines

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

  • If the 40 percent figure is taken relative to baselines that use linear extrapolation for future positions, a stronger learned trajectory predictor could narrow the margin; the paper does not report that head-to-head comparison on identical pseudo-labels.
  • The label-free claim is really "label-free at training time given LiDAR during development"; any sensor suite without reliable LiDAR coverage near ground clutter or beyond 100 meters would break the pseudo-label generator before the predictor is trained.
  • A natural stress test would be to run the same motion-alignment loss on a scene with a textureless sky, where ORB keypoints are scarce; the paper's own failure analysis suggests this is where the alignment supervision would be weakest.
  • Applying the pipeline to a dataset with dense ground-truth 3D boxes, instead of using MMAUD's survey-grade ground truth only for evaluation, would quantify how much pseudo-label noise the downstream stages tolerate.
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

3 major / 4 minor

Summary. The paper proposes an unsupervised/self-supervised pipeline for 3D UAV trajectory prediction from RGB and event cues. It first extracts motion-consistent trajectories from LiDAR/radar point clouds using a Temporal-KNN gradient filter, projects and aligns these trajectories into the image plane via ORB-based motion consistency to generate pseudo-labels, and then trains a Vision-Mamba detector with additional kinematic regression stages to forecast future 3D positions. Experiments on the MMAUD dataset compare current pose estimation and future trajectory prediction errors against several supervised and self-supervised image-only, audio-only, and audio-visual baselines.

Significance. The intended contribution is valuable: a label-free long-horizon prediction pipeline evaluated on a real-world multimodal benchmark, with a stated commitment to release code. The paper also addresses the rarely studied issue of sensor asynchrony and uses a conservative simulated-event setup. However, the headline claim of a roughly 40% error reduction without manual labels is not supported by the paper's own Table II; the label-free row improves over the best audio-visual baseline by only about 1.4%. Without error bars, multiple seeds, or a sensitivity analysis of the pseudo-labeling parameters, even this small margin is unverified. As written, the paper's central claim is substantially overstated, although the underlying pipeline may merit further study if the numerical evidence is strengthened.

major comments (3)
  1. [Abstract; §IV-D, Table II] The abstract claims 'reducing 5-second 3D error by around 40 percent without using any manual 3D labels.' This is contradicted by Table II: the label-free 'Ours – Proposed' row gives E5s = 6.20 m, which is only 1.4% below the best supervised audio-visual baseline AV-FDTI (E5s = 6.29 m). The 3.85 m value that yields a roughly 40% reduction belongs to the 'Ours – TLT' row, which uses manual terrestrial-laser-tracking labels and therefore cannot support the label-free claim. Section IV-D's stated comparison against AV-FDTI (6.20 vs. 6.29 m) is the only admissible evidence for the abstract's claim, and it is far weaker than advertised.
  2. [§IV-D, Tables II and III] All reported evaluations are single point estimates. No standard deviations, number of runs or seeds, or statistical significance tests are provided. The claimed label-free advantage at E5s is 0.09 m over AV-FDTI; without uncertainty quantification, that difference may be within run-to-run noise. The paper's central quantitative conclusions cannot be assessed from the reported numbers alone.
  3. [§III-B, Algorithm 1, Eq. (15); §V-B] The pseudo-label creation step depends on the fixed gradient threshold tau, and the alignment and DETR losses depend on weights lambda, lambda1, and lambda2. Section V-B explicitly concedes that no systematic sensitivity analysis was performed for these empirically chosen parameters. Because every downstream stage is trained from pseudo-labels produced by this filter, the claimed robustness of the label-free pipeline is not established; the edge-case failures acknowledged in Section V-A (occlusion, low altitude, and targets beyond 100 meters) are precisely the regimes where a fixed tau is most likely to fail. A sensitivity study over these parameters is needed to support the method's generality.
minor comments (4)
  1. [§III-A, Eq. (9)] Equation (9) defines the third component of X2 as J_t * alpha, but Equation (8) gives u_ddot = J_t * alpha + Jdot_t * V. The missing Jdot_t * V term should be corrected or the notation clarified.
  2. [§IV-D, Table I] Table I uses per-mille suffixes (e.g., '+77.8‰') where the surrounding text appears to mean percent improvement. Please make the units consistent throughout the table and its discussion.
  3. [§IV-B] The text states that closed-source baselines were excluded, yet VorasNet is described as fully re-implemented. Please clarify the exclusion criterion and provide a clear implementation or citation note. Several model names also contain spurious spaces ('V orasNet', 'A V-FDTI') and the text consistently renders 'UAV' as 'UA V'; these formatting issues should be fixed.
  4. [Title; §III-D; §IV-A] The title and abstract emphasize event cues, but the experiments use simulated events derived from RGB frames. Although this is disclosed in Sections III-D and IV-A, the abstract and title should make the simulated-event setup explicit so that readers do not infer evaluation with a real event sensor.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: pseudo-labels come from LiDAR/radar point clouds and ORB motion cues, while 5-second errors are measured against external Leica MS60 ground truth.

full rationale

The training loop is not circular. Temporal-KNN labels are extracted from raw LiDAR/radar point clouds (Algorithm 1, Eq. 10-15), not from the RGB/event network. ORB keypoint flow provides an independent image-space reference for cross-modal alignment (Algorithm 2, Eq. 17-21). The downstream Vision-Mamba model is trained against these externally generated pseudo-labels, and the reported 5-second RMSE values are evaluated against Leica MS60 TLT ground truth on the MMAUD dataset. The future-prediction head is trained with MSE against pseudo-ground-truth 3D positions derived from the same LiDAR-extracted trajectories, but the evaluation numbers are not reconstructible from those pseudo-labels by construction: the external TLT benchmark is independent of the training signal. The authors' self-citations (e.g., [14], [16], [17], [18], [28]) appear as dataset, background, or baseline references and are not load-bearing; the Temporal-KNN and alignment algorithms are fully specified in the paper. The abstract's 'around 40%' reduction is a numerical-claim concern (Table II: Ours-TLT E5s=3.85 vs Ours-Proposed E5s=6.20, with the best supervised baseline AV-FDTI at 6.29), and Section V-B concedes no sensitivity analysis for tau and other parameters, but these are correctness and robustness limitations, not circular reductions. No equation reduces to another by definition, and no fitted parameter is renamed as a prediction. Therefore the circularity score is 0.

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

The system is built from standard computer-vision machinery plus a set of empirically chosen thresholds and loss weights. The key undefended inputs are the temporal-smoothness assumption in Temporal-KNN, the use of ORB keypoints as alignment references, and the equivalence of simulated and real event cues. No new physical entities are introduced.

free parameters (5)
  • Neighbor count K = not reported
    Number of nearest neighbors in Temporal-KNN (Algorithm 1, Eq. 11); chosen empirically and not reported in the paper.
  • Temporal offset Delta t = not reported
    Frame offset used to build temporal vectors (Eq. 12-14); no value or sensitivity analysis is given.
  • Gradient threshold tau = not reported
    Threshold controlling temporal smoothness in Algorithm 1; Section V-B admits no systematic sensitivity analysis. It directly controls which points are kept as drone trajectories.
  • Alignment loss weight lambda = not reported
    Weight balancing cosine similarity and full-state Lstate in Algorithm 2 and Eq. (21); 'selected empirically based on validation performance' in Section III-D.
  • DETR loss weights lambda1, lambda2 = not reported
    Weights on L1 and GIoU losses in Eq. (31), selected empirically; standard but not specified.
assumptions (5)
  • domain assumption Moving objects maintain temporal vector consistency across unsynchronized LiDAR and radar frames, while static background vectors are random.
    Section III-B: 'objects in motion maintain temporal consistency' is used to justify the gradient filter with threshold tau.
  • domain assumption ORB keypoint tracks reliably reflect the drone's image-space motion and can serve as reference for alignment.
    Section III-C, Algorithm 2: projected LiDAR vectors are matched to nearest ORB keypoints and compared by cosine and full-state loss.
  • domain assumption Simulated events computed as pixel intensity differences between consecutive RGB frames approximate real event camera signals well enough for training.
    Section III-D and Limitations V-B; the authors call this a conservative lower bound, but no real event comparison is made.
  • standard math The unified fisheye projection model with parameters xi, k1, k2, fx, fy, cx, cy is known and differentiable.
    Equations (2)-(6) use the unified imaging model [31] for projections; correctness of the model and its intrinsics is assumed.
  • domain assumption The target drone is the only smoothly moving object in the point cloud segment, so Temporal-KNN isolates it rather than other moving objects.
    Algorithm 1 filters by temporal gradient magnitude; if other moving objects have similar smooth motion, they would also be retained. Not explicitly discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Label-Free Long-Horizon 3D UAV Trajectory Prediction via Motion-Aligned RGB and Event Cues." pith.science (2026). https://pith.science/paper/C2LPOD77

@misc{pith2026250703365,
  author       = {Pith},
  title        = {Pith review of: Label-Free Long-Horizon 3D UAV Trajectory Prediction via Motion-Aligned RGB and Event Cues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2LPOD77}},
  note         = {Machine review of arXiv:2507.03365}
}
read the original abstract

The widespread use of consumer drones has introduced serious challenges for airspace security and public safety. Their high agility and unpredictable motion make drones difficult to track and intercept. While existing methods focus on detecting current positions, many counter-drone strategies rely on forecasting future trajectories and thus require more than reactive detection to be effective. To address this critical gap, we propose an unsupervised vision-based method for predicting the three-dimensional trajectories of drones. Our approach first uses an unsupervised technique to extract drone trajectories from raw LiDAR point clouds, then aligns these trajectories with camera images through motion consistency to generate reliable pseudo-labels. We then combine kinematic estimation with a visual Mamba neural network in a self-supervised manner to predict future drone trajectories. We evaluate our method on the challenging MMAUD dataset, including the V2 sequences that feature wide-field-of-view multimodal sensors and dynamic UAV motion in urban scenes. Extensive experiments show that our framework outperforms supervised image-only and audio-visual baselines in long-horizon trajectory prediction, reducing 5-second 3D error by around 40 percent without using any manual 3D labels. The proposed system offers a cost-effective, scalable alternative for real-time counter-drone deployment. All code will be released upon acceptance to support reproducible research in the robotics community.

Figures

Figures reproduced from arXiv: 2507.03365 by the authors.

Figure 1
Figure 1. Core Motivation: Enabling Lead-Time-Aware UAV Defense. Most existing methods estimate the UAV’s current position (red dotted path), offer￾ing insufficient lead time to activate physical or autonomous countermeasures. In contrast, our method forecasts the UAV’s future trajectory (green dotted path) based on RGB and event cues, enabling timely response across a range of defense strategies. As illustrated below, differ… view at source ↗
Figure 2
Figure 2. Illustration of the heterogeneous sensing setup and the lack of [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Unsupervised temporal filtering via KNN vector gradient consistency. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of cross-modal motion alignment for pseudo-label refinement. Raw LiDAR and radar trajectories are projected into the fisheye image frame, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of direct 3D projection versus motion-aligned supervision. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Overview of the proposed label-free long-horizon self-supervised UAV trajectory prediction framework. The offline unsupervised branch extracts [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison of future trajectory prediction under dynamic constraints across four drone platforms (Part types illustration of Datasets). Top row shows [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: From limited FOV to robust tracking: a comparison of sensing [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Examples of real-world sensing challenges in the MMAUD dataset. (Left two) LiDAR scans suffer from cloudless sunlight interference, introducing significant point cloud noise that reduces spatial reliability. (Third) Audio channels are affected by continuous heavy machi…
Figure 10
Figure 10. Figure 10: Impact of Motion Dynamics on Pseudo-Label Error and the Effect of 2D Projection. We analyze how pseudo-label error correlates with UAV motion dynamics across velocity and acceleration domains. In both cases, 2D image-space projection significantly reduces the mean err…
Figure 11
Figure 11. Figure 11: Visual illustration of two edge-case scenarios that challenge our [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Drone Detection with Event Cameras

    cs.CV 2025-08 conditional novelty 2.0 of 10

    A survey of event camera-based drone detection that maps methods by data representation and covers tracking, forecasting, and propeller signature analysis.

Reference graph

Works this paper leans on

50 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [16]

    Unsupervised uav 3d trajectories estimation with sparse point clouds,

    H. Liang, Y . Yang, J. Hu, J. Yang, F. Liu, and S. Yuan, “Unsupervised uav 3d trajectories estimation with sparse point clouds,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

    work page 2025

  2. [1]

    Design of anti-drone laser weapon systems,

    S. Zhao, R. Xie, and J. Wan, “Design of anti-drone laser weapon systems,” in High-Power Lasers and Applications XI, vol. 11544. SPIE, 2020, pp. 24–30

    work page 2020

  3. [2]

    A 7.62 mm energetic bullet filled with ptfe-mg-based reactive materials for anti- drone application,

    Y . Cai, X. Feng, C. He, S. Zhang, S. Li, and J. Liu, “A 7.62 mm energetic bullet filled with ptfe-mg-based reactive materials for anti- drone application,” Journal of Materials Research and Technology , vol. 30, pp. 8749–8759, 2024

    work page 2024

  4. [3]

    Autonomous 3d moving target encirclement and interception with range measurement,

    F. Liu, S. Yuan, T.-M. Nguyen, and R. Su, “Autonomous 3d moving target encirclement and interception with range measurement,” in Pro- ceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , Hangzhou, China, October 2025

    work page 2025

  5. [4]

    On onboard lidar-based flying object detection,

    M. Vrba, V . Walter, V . Pritzl, M. Pliska, T. B ´aˇca, V . Spurn`y, D. He ˇrt, and M. Saska, “On onboard lidar-based flying object detection,” IEEE Transactions on Robotics , 2024

    work page 2024

  6. [5]

    Neptune: Nonen- tangling trajectory planning for multiple tethered unmanned vehicles,

    M. Cao, K. Cao, S. Yuan, T.-M. Nguyen, and L. Xie, “Neptune: Nonen- tangling trajectory planning for multiple tethered unmanned vehicles,” IEEE Transactions on Robotics , vol. 39, no. 4, pp. 2786–2804, 2023

    work page 2023

  7. [6]

    Action correction-enhanced multi- agent reinforcement learning for path planning in urban environments,

    H. Pan, L. Han, J. Yan, and R. Liu, “Action correction-enhanced multi- agent reinforcement learning for path planning in urban environments,” Unmanned Systems, 2025

    work page 2025

  8. [7]

    Distributed multi-robot sweep coverage for a region with unknown workload distribution,

    M. Cao, K. Cao, X. Li, S. Yuan, Y . Lyu, T.-M. Nguyen, and L. Xie, “Distributed multi-robot sweep coverage for a region with unknown workload distribution,” Autonomous Intelligent Systems , vol. 1, no. 1, p. 13, 2021

    work page 2021

Show all 50 references
  1. [8]

    Airswarm: Enabling cost-effective multi-uav research with cots drones,

    X. Li, K. Xu, F. Liu, R. Bai, S. Yuan, and L. Xie, “Airswarm: Enabling cost-effective multi-uav research with cots drones,” inProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025

    2025

  2. [9]

    Safety-assured high-speed navigation for mavs,

    Y . Ren, F. Zhu, G. Lu, Y . Cai, L. Yin, F. Kong, J. Lin, N. Chen, and F. Zhang, “Safety-assured high-speed navigation for mavs,” Science Robotics, vol. 10, no. 98, p. eado6187, 2025

    2025

  3. [10]

    Swarm of micro flying robots in the wild,

    X. Zhou, X. Wen, Z. Wang, Y . Gao, H. Li, Q. Wang, T. Yang, H. Lu, Y . Cao, C. Xuet al., “Swarm of micro flying robots in the wild,” Science Robotics, vol. 7, no. 66, p. eabm5954, 2022

    2022

  4. [11]

    Privacy’s sky-high battle: The use of unmanned aircraft systems for law enforcement in the european union,

    E. ¨O. Kurtpınar, “Privacy’s sky-high battle: The use of unmanned aircraft systems for law enforcement in the european union,” Journal of Intelligent & Robotic Systems , vol. 110, no. 3, p. 99, 2024

    2024

  5. [12]

    Fast identification and detec- tion algorithm for maneuverable unmanned aircraft based on multimodal data fusion,

    T. Luan, S. Zhou, Y . Zhang, and W. Pan, “Fast identification and detec- tion algorithm for maneuverable unmanned aircraft based on multimodal data fusion,” Mathematics, vol. 13, no. 11, p. 1825, 2025

    2025

  6. [13]

    Narco drones: tracing the evolution of cartel aerial tactics in mexico’s low-intensity conflicts,

    G. Krame, V . Vivoda, and A. Davies, “Narco drones: tracing the evolution of cartel aerial tactics in mexico’s low-intensity conflicts,” Small Wars & Insurgencies , vol. 34, no. 6, pp. 1095–1129, 2023

    2023

  7. [14]

    Mmaud: A comprehensive multi-modal anti-uav dataset for modern miniature drone threats,

    S. Yuan, Y . Yang, T. H. Nguyen, T.-M. Nguyen, J. Yang, F. Liu, J. Li, H. Wang, and L. Xie, “Mmaud: A comprehensive multi-modal anti-uav dataset for modern miniature drone threats,” in Proceedings of the 2024 IEEE International Conference on Robotics and Automation (ICRA) . Yo...

    2024

  8. [15]

    A comprehensive survey of unmanned aerial vehicles detection and classification using machine learning approach: Challenges, solutions, and future directions,

    M. H. Rahman, M. A. S. Sejan, M. A. Aziz, R. Tabassum, J.-I. Baik, and H.-K. Song, “A comprehensive survey of unmanned aerial vehicles detection and classification using machine learning approach: Challenges, solutions, and future directions,” Remote Sensing , vol. 16, no. 5, ...

    2024

  9. [17]

    Av-fdti: Audio-visual fusion for drone threat identification,

    Y . Yang, S. Yuan, J. Yang, T. H. Nguyen, M. Cao, T.-M. Nguyen, H. Wang, and L. Xie, “Av-fdti: Audio-visual fusion for drone threat identification,” Journal of Automation and Intelligence , vol. 3, no. 3, pp. 144–151, 2024

    2024

  10. [18]

    Audio array-based 3d uav trajectory estimation with lidar pseudo-labeling,

    A. H.-X. Lei, T. Deng, H. Wang, J. Yang, and S. Yuan, “Audio array-based 3d uav trajectory estimation with lidar pseudo-labeling,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

    2025

  11. [19]

    A new stereo fisheye event camera for fast drone detection and tracking,

    D. R. Da Costa, M. Robic, P. Vasseur, and F. Morbidi, “A new stereo fisheye event camera for fast drone detection and tracking,” in IEEE International Conference on Robotics and Automation , 2025

    2025

  12. [20]

    Evdetmav: Generalized mav detection from moving event cameras,

    Y . Zhang, Z. Ning, X. Zhang, S. Guo, P. Liu, and S. Zhao, “Evdetmav: Generalized mav detection from moving event cameras,” arXiv preprint arXiv:2506.19416, 2025

    2025 arXiv

  13. [21]

    Direct: A differential dynamic programming based framework for trajectory generation,

    K. Cao, M. Cao, S. Yuan, and L. Xie, “Direct: A differential dynamic programming based framework for trajectory generation,”IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 2439–2446, 2022

    2022

  14. [22]

    Anti-uav: A large-scale benchmark for vision-based uav tracking,

    N. Jiang, K. Wang, X. Peng, X. Yu, Q. Wang, J. Xing, G. Li, G. Guo, Q. Ye, J. Jiao et al., “Anti-uav: A large-scale benchmark for vision-based uav tracking,” IEEE Transactions on Multimedia , vol. 25, pp. 486–500, 2021

    2021

  15. [23]

    Vision-based anti-uav detection and tracking,

    J. Zhao, J. Zhang, D. Li, and D. Wang, “Vision-based anti-uav detection and tracking,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 12, pp. 25 323–25 334, 2022

    2022

  16. [24]

    Review and analysis of rgbt single object tracking methods: A fusion perspective,

    Z. Zhang, J. Wang, Z. Zang, L. Jin, S. Li, H. Wu, J. Zhao, and Z. Bo, “Review and analysis of rgbt single object tracking methods: A fusion perspective,” T-OMM, 2023

    2023

  17. [25]

    Etap: Event-based tracking of any point,

    F. Hamann, D. Gehrig, F. Febryanto, K. Daniilidis, and G. Gallego, “Etap: Event-based tracking of any point,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 186– 27 196

    2025

  18. [26]

    Motion and structure from event-based normal flow,

    Z. Ren, B. Liao, D. Kong, J. Li, P. Liu, L. Kneip, G. Gallego, and Y . Zhou, “Motion and structure from event-based normal flow,” in European Conference on Computer Vision . Springer, 2024, pp. 108– 125

    2024

  19. [27]

    Reducing the sim-to-real gap for event 15 cameras,

    T. Stoffregen, C. Scheerlinck, D. Scaramuzza, T. Drummond, N. Barnes, L. Kleeman, and R. Mahony, “Reducing the sim-to-real gap for event 15 cameras,” in Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16 . Springe...

    2020

  20. [28]

    Av-pedaware: Self- supervised audio-visual fusion for dynamic pedestrian awareness,

    Y . Yang, S. Yuan, M. Cao, J. Yang, and L. Xie, “Av-pedaware: Self- supervised audio-visual fusion for dynamic pedestrian awareness,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 1871–1877

    2023

  21. [29]

    A survey on event- based optical marker systems,

    N. J. Tofighi, M. Robic, F. Morbidi, and P. Vasseur, “A survey on event- based optical marker systems,” arXiv preprint arXiv:2504.20736 , 2025

    2025 arXiv

  22. [30]

    Event-aware distilled detr for object detection in an automotive con- text,

    D. Rossi, P. Vasseur, F. Morbidi, C. Demonceaux, and F. Rameau, “Event-aware distilled detr for object detection in an automotive con- text,” in IEEE INTELLIGENT VEHICLES SYMPOSIUM , 2025

    2025

  23. [31]

    3d to 2d bijection for spherical objects under equidistant fisheye projection,

    A. Ahmad, J. Xavier, J. Santos-Victor, and P. Lima, “3d to 2d bijection for spherical objects under equidistant fisheye projection,” Computer Vision and Image Understanding , vol. 125, pp. 172–183, 2014

    2014

  24. [32]

    Rein- forcement learning meets visual odometry,

    N. Messikommer, G. Cioffi, M. Gehrig, and D. Scaramuzza, “Rein- forcement learning meets visual odometry,” in European Conference on Computer Vision. Springer, 2024, pp. 76–92

    2024

  25. [33]

    Gradient-based hyper- parameter optimization through reversible learning,

    D. Maclaurin, D. Duvenaud, and R. Adams, “Gradient-based hyper- parameter optimization through reversible learning,” in International conference on machine learning . PMLR, 2015, pp. 2113–2122

    2015

  26. [34]

    Yolov8: A novel object detection algo- rithm with enhanced performance and robustness,

    R. Varghese and M. Sambath, “Yolov8: A novel object detection algo- rithm with enhanced performance and robustness,” in 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS). IEEE, 2024, pp. 1–6

    2024

  27. [35]

    Yolov10: Real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han et al. , “Yolov10: Real-time end-to-end object detection,” Advances in Neural Information Processing Systems, vol. 37, pp. 107 984–108 011, 2024

    2024

  28. [36]

    Dronechase: A mobile and automated cross-modality system for continuous drone tracking,

    N. V ora, Y . Wu, J. Liu, and P. Nguyen, “Dronechase: A mobile and automated cross-modality system for continuous drone tracking,” in Proceedings of the ninth workshop on micro aerial vehicle networks, systems, and applications , 2023, pp. 8–14

    2023

  29. [37]

    Visualnet: An end-to-end human visual system inspired framework to reduce inference latency of deep neural networks,

    T. Wang, J. Zhang, J. Xiong, S. Bian, Z. Yan, M. Huang, J. Zhuang, T. Sato, X. Xu, and Y . Shi, “Visualnet: An end-to-end human visual system inspired framework to reduce inference latency of deep neural networks,” IEEE Transactions on Computers, vol. 71, no. 11, pp. 2717– 2727, 2022

    2022

  30. [38]

    Is someone speaking? exploring long-term temporal features for audio- visual active speaker detection,

    R. Tao, Z. Pan, R. K. Das, X. Qian, M. Z. Shou, and H. Li, “Is someone speaking? exploring long-term temporal features for audio- visual active speaker detection,” in Proceedings of the 29th ACM international conference on multimedia , 2021, pp. 3927–3935

    2021

  31. [39]

    Yolov4: Op- timal speed and accuracy of object detection,

    A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “Yolov4: Op- timal speed and accuracy of object detection,” arXiv preprint arXiv:2004.10934, 2020

    2004 arXiv

  32. [40]

    Keypoint-guided ef- ficient pose estimation and domain adaptation for micro aerial vehicles,

    Y . Zheng, C. Zheng, J. Shen, P. Liu, and S. Zhao, “Keypoint-guided ef- ficient pose estimation and domain adaptation for micro aerial vehicles,” IEEE Transactions on Robotics , 2024

    2024

  33. [41]

    Drone detection and tracking system based on fused acoustical and optical approaches,

    S. Ding, X. Guo, T. Peng, X. Huang, and X. Hong, “Drone detection and tracking system based on fused acoustical and optical approaches,” Advanced Intelligent Systems , vol. 5, no. 10, p. 2300251, 2023

    2023

  34. [42]

    A unified transformer based tracker for anti-uav tracking,

    Q. Yu, Y . Ma, J. He, D. Yang, and T. Zhang, “A unified transformer based tracker for anti-uav tracking,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 3036–3046

    2023

  35. [43]

    Multimodal future localization and emergence prediction for objects in egocentric view with a reachability prior,

    O. Makansi, O. Cicek, K. Buchicchio, and T. Brox, “Multimodal future localization and emergence prediction for objects in egocentric view with a reachability prior,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 4354–4363

    2020

  36. [44]

    Robust collaborative discrim- inative learning for rgb-infrared tracking,

    X. Lan, M. Ye, S. Zhang, and P. Yuen, “Robust collaborative discrim- inative learning for rgb-infrared tracking,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 32, no. 1, 2018

    2018

  37. [45]

    Adaptive and background- aware vision transformer for real-time uav tracking,

    S. Li, Y . Yang, D. Zeng, and X. Wang, “Adaptive and background- aware vision transformer for real-time uav tracking,” in Proceedings of the IEEE/CVF international conference on computer vision , 2023, pp. 13 989–14 000

    2023

  38. [46]

    Learning short-term spatial- temporal dependency for uav 2d trajectory forecasting,

    Z. Siyuan, L. Yang, X. Liu, and L. Wang, “Learning short-term spatial- temporal dependency for uav 2d trajectory forecasting,” IEEE Sensors Journal, 2024

    2024

  39. [47]

    Continuous-time state estimation methods in robotics: A survey,

    W. Talbot, J. Nubert, T. Tuna, C. Cadena, F. D ¨umbgen, J. Tordesillas, T. D. Barfoot, and M. Hutter, “Continuous-time state estimation methods in robotics: A survey,” arXiv preprint arXiv:2411.03951 , 2024

    2024 arXiv

  40. [48]

    An efficient b-spline-based kinodynamic replanning framework for quadrotors,

    W. Ding, W. Gao, K. Wang, and S. Shen, “An efficient b-spline-based kinodynamic replanning framework for quadrotors,” IEEE Transactions on Robotics, vol. 35, no. 6, pp. 1287–1306, 2019

    2019

  41. [49]

    Geometrically constrained tra- jectory optimization for multicopters,

    Z. Wang, X. Zhou, C. Xu, and F. Gao, “Geometrically constrained tra- jectory optimization for multicopters,” IEEE Transactions on Robotics , vol. 38, no. 5, pp. 3259–3278, 2022

    2022

  42. [50]

    Incorporating control inputs in continuous- time gaussian process state estimation for robotics,

    S. Lilge and T. D. Barfoot, “Incorporating control inputs in continuous- time gaussian process state estimation for robotics,” Robotica, vol. 43, no. 3, pp. 1067–1086, 2025

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.