Pith. sign in

REVIEW 5 major objections 6 minor 63 references

Eyes on the Road, Mind Beyond Vision: Context-Aware Multi-modal Enhanced Risk Anticipation

T0 review · 5 major / 6 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 CAMERA fuses dashcam video, driver gaze, and scene text to anticipate accidents earlier while reducing false alarms.

desk verdict A coherent fusion architecture with plausible accuracy numbers, but the paper's central adaptive-threshold false-alarm claim is never actually measured. read the letter →

arxiv 2507.06444 v2 pith:A5WO7V5R submitted 2025-07-08 cs.CE

Jiaxun Zhang , Haicheng Liao , Yumu Xie , Chengyue Wang , Yanchen Guan , Bin Rao , Zhenning Li This is my paper
classification cs.CE
keywords trafficaccidentanticipationdriverattentionmodelingmulti-modalfusionadaptivethresholdvision-languagereasoningDADA-2000dashcamvideo
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

This paper claims that accident anticipation improves when dashcam video, textual scene descriptions, and driver attention maps are fused in one pipeline with a learned threshold that adapts to scene complexity. The proposed CAMERA system reports state-of-the-art results on DADA-2000: 80.52% average precision, 89.64% AUC, 4.06 seconds time-to-accident at 50% recall, and 4.48 seconds mean time-to-accident. It also reports gains on CCD, DAD, and A3D, including longer lead times and lower false-positive rates than fixed-threshold systems. If these results hold, they support a design principle for driver-assistance systems: warnings should depend not only on predicted risk but on how complex and how gaze-demanding the current scene is.

What carries the argument

The load-bearing mechanism is the dynamic threshold $\tau_t = 0.5 + \lambda_1 E(X_f^{(t)}) - \lambda_2 \|F_{\mathrm{context}}^{(t)}\|_2$ (Eq. 9): it raises or lowers the alert cutoff using driver attention entropy $E$ and the norm of textual context features, keeping the threshold inside [0.3, 0.7]. This is what the paper credits for suppressing false alarms while preserving recall. Around it, the other components—multi-scale attention on gaze maps, hierarchical cross-modal fusion (HSFA, PCN, CoAT, BiBA), the Bi-GRU with adaptive memory blending, and the Geo-Context VLM for 3D-grounded alerts—supply the features that the adaptive threshold acts on.

What would settle it

Re-run DRIVE, LOTVS-CAP, GSC, CRASH, and W3AL under CAMERA's exact splits, sampling windows, and preprocessing; if the reported advantages on DAD (+8.6% AP) and A3D (+2.12s mTTA over GSC) shrink or reverse, the generalization claim fails.

Watch

Extended reading notes

Core claim

CAMERA's central discovery is that treating the risk threshold as part of the learned model, rather than as a fixed constant, yields both higher accuracy and earlier warnings. The threshold in Eq. (9) moves within [0.3, 0.7] and is lowered when driver gaze entropy is high and scene complexity is low, so benign environments do not trigger false alarms while genuinely evolving hazards are flagged sooner. The framework couples this adaptive threshold with a hierarchical fusion of RGB frames, BERT text embeddings, and gaze heatmaps, a bidirectional GRU over a 3-5 second anticipation window, and a Geo-Context VLM that converts numerical spatial predictions into readable alerts. On DADA-2000 the method reports 80.52% AP and 89.64% AUC, above DRIVE and LOTVS-CAP, and on the three cross-dataset benchmarks it reports the best or second-best AP and mTTA in every column.

Load-bearing premise

The cross-dataset gains assume the published baseline numbers for CCD, DAD, and A3D were produced under the same train/test splits, sampling windows, and preprocessing as CAMERA; the paper does not rerun those baselines in its own harness.

Editorial extensions

If this is right

  • If CAMERA's results transfer, ADAS warning systems can replace fixed confidence cutoffs with scene-adaptive thresholds, cutting nuisance alerts in low-complexity driving.
  • The reported 4.06-4.48 second lead times exceed typical human reaction windows of 3-5 seconds, suggesting the fused pipeline can provide actionable warning before a crash event.
  • The Geo-Context VLM alerts, such as "pedestrian 2.1m ahead in left blind spot," convert probabilistic risk into spatial language a driver can act on, addressing interpretability of anticipation models.
  • On limited training data and simulated sensor dropout, CAMERA retains higher AP and lead time than DRIVE and LOTVS-CAP, indicating robustness to partial observability in deployment.

Reading between the lines

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

  • The adaptive threshold's exact form resembles a homeostatic balance between attentional load and environmental complexity; a natural extension is to test whether gaze entropy alone, without text embeddings, can drive the threshold in datasets that lack textual annotations.
  • The cross-dataset gains suggest that gaze maps might carry transferable cues about driver inattention that generalize across countries and traffic cultures; that hypothesis is testable by ablating the attention stream on DAD and A3D.
  • Because the Geo-Context VLM depends on Metric3Dv2 and WildCamera for metric depth and calibration, its alert accuracy is only as reliable as those upstream monocular estimates; an evaluation that compares alert positions against LiDAR ground truth would isolate this dependency.
  • The 24% safety margin quoted for the dynamic threshold in the qualitative section is computed from one illustrated scenario; a systematic margin analysis across all DADA-2000 videos would show whether the margin is a general property or a case-specific illustration.
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 / 6 minor

Summary. The manuscript proposes CAMERA, a multi-modal framework for traffic accident anticipation that fuses dashcam RGB frames, textual scene descriptions, and driver attention maps through a hierarchical fusion pipeline and a Bi-GRU temporal model, and introduces a scene-complexity-adaptive warning threshold (Eq. 9) plus a Geo-Context VLM for interpretable alerts. The paper reports state-of-the-art AP/AUC/TTA on DADA-2000 and improved results on CCD, DAD, and A3D, and claims that the adaptive threshold significantly reduces false positives while extending lead time. The experiments include ablations, limited-data robustness tests, and qualitative visualizations, but the central false-alarm claim is never measured and no code or error bars are provided.

Significance. If the reported results hold, CAMERA would offer a useful contribution to multimedia-based driver assistance: it integrates three modalities, provides interpretable spatially grounded alerts, and proposes a thresholding mechanism that adapts to scene complexity. The paper's strengths include evaluation on multiple public benchmarks, comparison with several recent baselines, and a systematic ablation of the main modules. However, the distinctive contribution, adaptive false-alarm reduction, is never directly measured, and the implementation contains notation errors that make the method description unreliable as written. The significance is therefore contingent on substantial verification.

major comments (5)
  1. [Abstract; Section 4.6] The claim that CAMERA "significantly reduces false positives" is not supported by any measured metric. AP and AUC are threshold-free ranking metrics, and TTA@R50 and mTTA condition on recall; none of them counts the number of non-accident frames that trigger a warning. Section 4.6 presents only hand-picked examples (a single non-accident scene and one conflict case), with no aggregate false-positive rate, false-alarms-per-hour, or precision-at-matched-recall. This is load-bearing because the adaptive threshold in Eq. (9) is the paper's main novel contribution.
  2. [Section 3.7, Eq. (9)] Equation (9) defines tau_t = 0.5 + lambda_1 E(X_f) - lambda_2 ||F_context||_2, with lambda_1 and lambda_2 in [0, 0.2], and the text states that these coefficients ensure tau_t in [0.3, 0.7]. The paper does not bound or normalize E(X_f) or ||F_context||_2, so the range guarantee does not follow from the stated constraints. The authors should either prove the bound, clip or normalize the terms, or report the empirical distribution of tau_t and the resulting false-alarm rates.
  3. [Section 4.4, Table 2] The cross-dataset comparisons on CCD, DAD, and A3D assume that the cited baseline numbers (Ustring, GSC, CRASH, W3AL) were produced under the same train/test splits, sampling windows, and preprocessing as CAMERA. No protocol alignment is described and the baselines are not rerun in the authors' harness. The claimed +8.6% AP on DAD and +2.12s mTTA advantage over GSC on A3D therefore may not survive a controlled comparison.
  4. [Section 3.5, Eq. (4)] Equation (4) contains corrupted notation: it states that mu denotes the Swin Transformer and that [·;·] represents graph convolution network, which are not meaningful definitions in a GRU update-gate equation. This makes the method description unreliable as written and must be corrected with the actual gating operations.
  5. [Section 4.5, Table 3] The data-efficiency narrative is contradicted by the numbers in Table 3. At 50% training data, CAMERA retains 92.1% of its full-data AP (0.7413/0.8052), while LOTVS-CAP retains 98.8% (0.7406/0.75), so by this retention measure CAMERA is less, not more, data-efficient. The text's claim that CAMERA is "the first model optimizing both performance maxima and data-efficiency minima" needs a different metric or a corrected analysis.
minor comments (6)
  1. [Section 4.6] The paper reports a "22% safety margin" in one paragraph and a "24% safety margin" in a later paragraph; these numbers should be reconciled.
  2. [Table 5 caption] The caption says "MFE donates Multi-Modal Feature Extraction" and "AHF donates Adaptive Hierarchical Fusion"; "donates" should be "denotes".
  3. [Section 3.6] There is a duplicated phrase "Consequently, Consequently" in the description of the Geo-Context VLM.
  4. [Section 4.3] The implementation details refer to "attention Sentropy"; this appears to be a typo for "attention entropy".
  5. [Figure 2] Figure 2 lists modules PCN and CoAT that are not described in the text; these acronyms should be defined or the figure should be aligned with the described architecture.
  6. [Section 4.2] The definitions of TTA@R50 and mTTA are terse; the paper should specify how a "correctly anticipated" accident is determined and how the time difference is measured.

Circularity Check

2 steps flagged · score 6.0 of 10

Adaptive-threshold benefits are read off Eq. 9 rather than measured; the AP/AUC claims remain externally benchmarked but the paper's central false-alarm-reduction claim reduces to the fitted threshold itself.

  1. self definitional [Section 3.7 (Eq. 9) and Section 4.6]
    "To reduce false alarms and increase robustness, the threshold τ_t dynamically modulates the prediction sensitivity while λ1, λ2∈[0,0.2] are learnable coefficients ensuring τ_t∈[0.3,0.7] ... τ_t = 0.5+λ1E(X_f)−λ2||F_context||_2. ... Quantitative analysis shows CAMERA triggers warnings when predicted probability (0.4) exceeds the dynamic threshold, achieving a 0.25-second advantage over static thresholds (0.5) while maintaining a 22% safety margin."

    The reported 0.25-second lead-time advantage is not evaluated against held-out false-alarm costs; it follows automatically from having a threshold below 0.5, since any monotonically rising risk curve crosses a lower threshold earlier. The '22% safety margin' is just (0.5−τ_t)/0.5, a direct rearrangement of the τ_t value produced by Eq. 9. Thus the headline benefit of the adaptive threshold is a property of the fitted threshold values, not an independently measured outcome, and no false-positive-rate or precision-at-recall metric appears anywhere in Section 4.2.

  2. fitted input called prediction [Section 4.6, Figure 4 analysis]
    "As validated in Fig 4, the divergence between accident probability (0.0–0.29) and complexity reduction (lowered to 0.38) confirms CAMERA's 24% safety margin below the static 0.5 benchmark."

    The '24% safety margin' is computed directly from the threshold value 0.38 that Eq. 9 outputs for that scene, i.e. (0.5−0.38)/0.5. The passage presents this arithmetic consequence of the learned threshold as independent confirmation that the mechanism suppresses false alarms. Because no held-out false-alarm rate is measured and the example is selected from the same regime used to fit the model, the claimed false-alarm reduction reduces to re-describing the fitted threshold rather than predicting it.

full rationale

The accuracy and anticipation-time claims on DADA-2000, CCD, DAD and A3D are compared with external or published baselines and do not, by themselves, reduce to the model's inputs; there is no load-bearing uniqueness theorem, no imported ansatz from the authors' prior work, and no renaming of a known result that carries the central derivation. However, the paper's distinctive contribution is the adaptive thresholding mechanism, and the quantitative support for that mechanism is circular in a localized but important way. Eq. 9 defines τ_t as 0.5+λ1E(X_f)−λ2||F_context||_2 with learned coefficients, and Section 4.6 then 'validates' the mechanism by citing the margin between 0.5 and the threshold values this same equation produces, plus a lead-time advantage that any lower threshold would produce by construction. The paper never reports a false-alarm rate, false-alarms-per-hour, or precision at matched recall, so the central assertion that the adaptive threshold 'significantly reduces false positives' is supported only by the fitted threshold's own output. This justifies a score of 6 rather than 0: the AP/AUC benchmarking is self-contained, but the adaptive-threshold prediction that is the paper's headline novelty reduces by construction to the fitted threshold. Separately, the claim that λ1,λ2∈[0,0.2] ensures τ_t∈[0.3,0.7] is mathematically unsubstantiated because E(X_f) and ||F_context||_2 are unbounded in the text; that is a correctness risk and is not counted as circularity here.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central mechanism of the paper is the adaptive threshold, which depends on two learned coefficients and two unvalidated proxy signals (text-embedding norm and attention entropy). The model also relies on external pretrained perception tools and assumes gaze and text are available at inference. No new physical entities are introduced.

free parameters (4)
  • lambda_1 = learnable, constrained to [0, 0.2]
    Coefficient for attention-entropy term E(X_f) in the adaptive threshold (Eq. 9). Fitted end-to-end on the training set; no independent validation of the resulting threshold behavior.
  • lambda_2 = learnable, constrained to [0, 0.2]
    Coefficient for scene-complexity term ||F_context||_2 in Eq. 9. Fitted end-to-end; the paper does not ablate different ranges or report learned values.
  • loss weights = 0.5 for KL, 0.1 for smoothness
    Hand-selected weights in Eq. 10; no sensitivity analysis is provided.
  • adaptive threshold bounds = [0.3, 0.7]
    Arbitrary hard bounds imposed on tau_t in Section 3.1; no justification for these limits.
assumptions (5)
  • domain assumption Textual descriptions of weather, road layout, and events are available at both training and test time
    Assumed in Section 3.3 and used in all experiments. DADA-2000 does not provide free-form text annotations for every clip; the paper does not specify how text is obtained or verified for the four datasets.
  • domain assumption High-resolution driver gaze/attention maps are available at inference time
    Eq. 1 and the whole MFE pathway condition predictions on attention heatmaps. The conclusion admits dependence on high-resolution gaze data, so the method cannot be deployed without a gaze tracker.
  • domain assumption Pre-trained Grounding DINO, Metric3Dv2, and WildCamera provide accurate object, depth, and camera-calibration cues
    Geo-Context VLM (Section 3.6) uses these external models; their output errors propagate directly into the generated alerts.
  • ad hoc to paper The L2 norm of the textual embedding is a valid measure of scene complexity
    Eq. 9 uses ||F_context||_2 as the complexity signal with no derivation or empirical validation that it correlates with traffic density or risk.
  • ad hoc to paper Attention-map entropy E(X_f) is a valid measure of cognitive load
    Used in Eq. 9 to modulate the threshold; no evidence links gaze entropy to accident risk in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Eyes on the Road, Mind Beyond Vision: Context-Aware Multi-modal Enhanced Risk Anticipation." pith.science (2026). https://pith.science/paper/A5WO7V5R

@misc{pith2026250706444,
  author       = {Pith},
  title        = {Pith review of: Eyes on the Road, Mind Beyond Vision: Context-Aware Multi-modal Enhanced Risk Anticipation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A5WO7V5R}},
  note         = {Machine review of arXiv:2507.06444}
}
read the original abstract

Accurate accident anticipation remains challenging when driver cognition and dynamic road conditions are underrepresented in predictive models. In this paper, we propose CAMERA (Context-Aware Multi-modal Enhanced Risk Anticipation), a multi-modal framework integrating dashcam video, textual annotations, and driver attention maps for robust accident anticipation. Unlike existing methods that rely on static or environment-centric thresholds, CAMERA employs an adaptive mechanism guided by scene complexity and gaze entropy, reducing false alarms while maintaining high recall in dynamic, multi-agent traffic scenarios. A hierarchical fusion pipeline with Bi-GRU (Bidirectional GRU) captures spatio-temporal dependencies, while a Geo-Context Vision-Language module translates 3D spatial relationships into interpretable, human-centric alerts. Evaluations on the DADA-2000 and benchmarks show that CAMERA achieves state-of-the-art performance, improving accuracy and lead time. These results demonstrate the effectiveness of modeling driver attention, contextual description, and adaptive risk thresholds to enable more reliable accident anticipation.

Figures

Figures reproduced from arXiv: 2507.06444 by the authors.

Figure 1
Figure 1. Overview of CAMERA’s anticipation process. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of our proposed CAMERA model, illustrating the pipeline from multi-modal inputs (RGB frames, scene context, driver attention) to risk anticipation and interpretable geo-context alerts. Modules include HSFA (Hierarchical Scale Fusion with Attention), PCN (Position-wise Correspondence Network), CoAT (Cross-Modal Co-Activation Transformer), BiBA (Bifurcated Basis Aggregation), and AMB (Adaptive Mem… view at source ↗
Figure 3
Figure 3. Comparison of Geo-Context VLM with LATTE-Flamingo[58], LLaVa-NEXT[33] and VTimeLLM[19] in different situation. 4.2 Evaluation Metrics To evaluate CAMERA’s accident anticipation performance, we adopt four widely used metrics: • Average Precision: AP quantifies the precision-recall curve by integrating precision values across all recall levels, with higher values indicating more reliable performance. • Area Under Curv… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: ). In the non-accident case ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: DADA-2000 conflict case: High-risk anticipation vs. null outcome, agents (yellow) and temporal data. thresholds (0.5) while maintaining a 22% safety margin. These re￾sults confirm CAMERA improves prognostic sensitivity, extending emergency response times and enabling a…
Figure 7
Figure 7. Figure 7: DADA-2000 conflict case: Low-risk anticipation vs. null outcome, agents (yellow) and temporal data. The dynamic threshold based on environmental complexity re￾duction derives from its ability to adapt to evolving risks. A static 0.5 threshold may over-alarm in low-traf…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 55 canonical work pages

  1. [1]

    Ismail Abdulrashid, Reza Zanjirani Farahani, Shamkhal Mammadov, Mohamed Khalafalla, and Wen-Chyuan Chiang. 2024. Explainable artificial intelligence in transport Logistics: Risk analysis for road accidents. Transportation Research Part E: Logistics and Transportation Review 186 (2024), 103563

    work page 2024

  2. [2]

    Taif Anjum, Daya Kumar, and Apurva Narayan. 2023. Spatio-temporal Anal- ysis of Dashboard Camera Videos for Time-To-Accident Forecasting. In 2023 International Joint Conference on Neural Networks (IJCNN) . IEEE, 1–8

    work page 2023

  3. [3]

    Mohammad BaniSalman, Mohammad Aljaidi, Najat Elgeberi, Ayoub Alsarhan, and Rabia Emhamed Al Mamlook. 2025. VRDeepSafety: A Scalable VR Simula- tion Platform with V2X Communication for Enhanced Accident Prediction in Autonomous Vehicles. World Electric Vehicle Journal 16, 2 (2025), 82

    work page 2025

  4. [4]

    Wentao Bao, Qi Yu, and Yu Kong. 2020. Uncertainty-based traffic accident anticipation with spatio-temporal relational learning. In Proceedings of the 28th ACM International Conference on Multimedia . 2682–2690

    work page 2020

  5. [5]

    Wentao Bao, Qi Yu, and Yu Kong. 2021. Drive: Deep reinforced accident an- ticipation with visual explanation. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 7619–7628

    work page 2021

  6. [6]

    Joe Beck, Ramin Arvin, Steve Lee, Asad Khattak, and Subhadeep Chakraborty

    work page

  7. [7]

    Xiangpeng Cai, Bowen Lv, Hanchen Yao, Ting Yang, and Houde Dai. 2025. V- FCW: Vector-based forward collision warning algorithm for curved road conflicts using V2X networks. Accident Analysis & Prevention 210 (2025), 107836

    work page 2025

  8. [8]

    Jingzheng Chai, Jianting Liu, Jingluo Huang, and Chunyan Huang. 2025. GACNet: Interactive Prediction of Surrounding Vehicles Behavior under High Collision Risk. Advanced Intelligent Systems (2025), 2401040

    work page 2025

Show all 63 references
  1. [9]

    Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, and Min Sun. 2017. Anticipating acci- dents in dashcam videos. In Computer Vision–ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13. Springer, 136–153

    2017

  2. [10]

    Nobhonil Roy Choudhury, Sreeja Bhattacharjee, Saptarsi Ghosh, Shivnath Ghosh, and Pranashi Chakraborty. 2025. Intelligent Traffic Management and Accident Prevention System with Vehicle Counting and Distance-Based Brake Control. In Driving Green Transportation System Through Ar...

    2025

  3. [11]

    Chowdhury, R

    N. Chowdhury, R. Patel, and V. Kumar. 2023. Flamingo: A lightweight visual- language model for real-time applications. Journal of Artificial Intelligence Re- search 58, 2 (2023), 111–129

    2023

  4. [12]

    Rosa Pia Devanna, Giulio Reina, Fernando Auat Cheein, and Annalisa Milella

  5. [13]

    Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, and Hongkai Yu. 2021. DADA: Driver attention prediction in driving accident scenarios. IEEE transac- tions on intelligent transportation systems 23, 6 (2021), 4959–4971

    2021

  6. [14]

    Langxiong Gan, Ziyi Gao, Xiyu Zhang, Yi Xu, Ryan Wen Liu, Cheng Xie, and Yaqing Shu. 2025. Graph neural networks enabled accident causation prediction for maritime vessel traffic. Reliability Engineering & System Safety 257 (2025), 110804

    2025

  7. [15]

    Jake Grigsby, Yuke Zhu, Michael S Ryoo, and Juan Carlos Niebles. [n. d.]. VLM Q-Learning: Aligning Vision-Language Models for Interactive Decision-Making. In Scaling Self-Improving Foundation Models without Human Supervision

  8. [16]

    Muhua Guan, Bo Yang, Zheng Wang, Chenchang Li, and Kimihiko Nakano. 2025. An Experimental Study on Drivers’ Eye Movement Behavior When Using an Automated Lane Change System. IEEE Transactions on Intelligent Transportation Systems (2025)

    2025

  9. [17]

    Ziang Guo, Zakhar Yagudin, Artem Lykov, Mikhail Konenkov, and Dzmitry Tsetserukou. 2024. VLM-Auto: VLM-based Autonomous Driving Assistant with Human-like Behavior and Understanding for Complex Road Scenes. In 2024 2nd International Conference on Foundation and Large Language M...

    2024

  10. [18]

    Seamie Hayes, Ganesh Sistu, and Ciaran Eising. 2025. Revisiting Birds Eye View Perception Models with Frozen Foundation Models: DINOv2 and Metric3Dv2. arXiv preprint arXiv:2501.08118 (2025)

    2025 arXiv

  11. [19]

    Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. 2024. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 14271–14280

    2024

  12. [20]

    Rongtian Huo, Junkang Chen, Ye Zhang, and Qing Gao. 2025. 3D skeleton aware driver behavior recognition framework for autonomous driving system. Neurocomputing 613 (2025), 128743

    2025

  13. [21]

    Zahid Hussain, Shabna Sayed Mohammed, Charitha Dias, Qinaat Hussain, and Wael KM Alhajyaseen. 2025. Empirical analysis of car-following behavior: Im- pacts of driver demographics, leading vehicle types, and speed limits on driver behavior and safety. Transportation Research Pa...

    2025

  14. [22]

    Sandesh Jain, Surendrabikram Thapa, Kuan-Ting Chen, A Lynn Abbott, and Abhijit Sarkar. 2024. Semantic understanding of traffic scenes with large vision language models. In 2024 IEEE Intelligent Vehicles Symposium (IV) . IEEE, 1580– 1587

    2024

  15. [23]

    Muhammad Monjurul Karim, Yu Li, Ruwen Qin, and Zhaozheng Yin. 2022. A Dynamic Spatial-Temporal Attention Network for Early Anticipation of Traffic Accidents. IEEE Transactions on Intelligent Transportation Systems 23, 7 (2022), 9590–9600. doi:10.1109/TITS.2022.3155613

    2022

  16. [24]

    Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. 2023. Segment anything in high quality. Advances in Neural Information Processing Systems 36 (2023), 29914–29934

    2023

  17. [25]

    Young Woo Kim and Sol Hee Yoon. 2025. Understanding drivers’ situation awareness in highly automated driving using SAGAT, SART, and eye-tracking data. Transportation Research Part F: Traffic Psychology and Behaviour 109 (2025), 1437–1450

    2025

  18. [26]

    Yuto Kumamoto, Kento Ohtani, Daiki Suzuki, Minori Yamataka, and Kazuya Takeda. 2025. AAT-DA: Accident Anticipation Transformer with Driver Attention. In Proceedings of the Winter Conference on Applications of Computer Vision . 1142– 1151

    2025

  19. [27]

    Jérémy Lachance-Tremblay, Zoubeir Tkiouat, Pierre-Majorique Léger, Ann- Frances Cameron, Ryad Titah, Constantinos K Coursaris, and Sylvain Sénécal

  20. [28]

    Lei-Lei Li, Jianwu Fang, and Jianru Xue. 2024. Cognitive Traffic Accident Antici- pation. IEEE Intelligent Transportation Systems Magazine (2024)

    2024

  21. [29]

    Mingyong Li, Xiaowu Zhu, Jiangtao Ji, Xin Jin, Baoming Li, Kaikang Chen, and Weipeng Zhang. 2025. Visual perception enabled agriculture intelligence: A selective seedling picking transplanting robot. Computers and Electronics in Agriculture 229 (2025), 109821

    2025

  22. [30]

    International Journal of Human-Computer Studies 193 (2025), 103366

    A gaze-based driver distraction countermeasure: Comparing effects of multimodal alerts on driver’s behavior and visual attention. International Journal of Human-Computer Studies 193 (2025), 103366

    2025

  23. [31]

    Haicheng Liao, Yongkang Li, Zhenning Li, Zilin Bian, Jaeyoung Lee, Zhiyong Cui, Guohui Zhang, and Chengzhong Xu. 2024. Real-time accident anticipation for autonomous driving through monocular depth-enhanced 3D modeling. Accident Analysis & Prevention 207 (2024), 107760

    2024

  24. [32]

    Haicheng Liao, Yongkang Li, Zhenning Li, Chengyue Wang, Zhiyong Cui, Shengbo Eben Li, and Chengzhong Xu. 2024. A Cognitive-Based Trajectory Prediction Approach for Autonomous Driving. IEEE Transactions on Intelligent Vehicles (2024)

    2024

  25. [33]

    Haicheng Liao, Hanlin Kong, Bonan Wang, Chengyue Wang, Wang Ye, Zheng- bing He, Chengzhong Xu, and Zhenning Li. 2025. Cot-drive: Efficient motion forecasting for autonomous driving with llms and chain-of-thought prompting. arXiv preprint arXiv:2503.07234 (2025)

    2025 arXiv

  26. [34]

    Haicheng Liao, Huanming Shen, Zhenning Li, Chengyue Wang, Guofa Li, Yiming Bie, and Chengzhong Xu. 2024. GPT-4 enhanced multimodal grounding for autonomous driving: Leveraging cross-modal attention with large language models. Communications in Transportation Research 4 (2024), 100116

    2024

  27. [35]

    Haicheng Liao, Haoyu Sun, Huanming Shen, Chengyue Wang, Chunlin Tian, KaHou Tam, Li Li, Chengzhong Xu, and Zhenning Li. 2024. CRASH: Crash Recog- nition and Anticipation System Harnessing with Context-Aware and Temporal Focus Attentions. In Proceedings of the 32nd ACM Internat...

    2024

  28. [36]

    Haicheng Liao, Yongkang Li, Chengyue Wang, Yanchen Guan, Kahou Tam, Chun- lin Tian, Li Li, Chengzhong Xu, and Zhenning Li. 2024. When, Where, and What? A Benchmark for Accident Anticipation and Localization with Large Language Models. In Proceedings of the 32nd ACM Internation...

    2024

  29. [37]

    Ciyun Lin, Shaoqi Zhang, Bowen Gong, and Hongchao Liu. 2024. Near-crash risk identification and evaluation for takeout delivery motorcycles using roadside LiDAR. Accident Analysis & Prevention 199 (2024), 107520

    2024

  30. [38]

    Qin Liu, Fei Wang, Chaowei Xiao, and Muhao Chen. 2025. VLM-Guard: Safe- guarding Vision-Language Models via Fulfilling Safety Alignment Gap. arXiv preprint arXiv:2502.10486 (2025)

    2025 arXiv

  31. [39]

    Haicheng Liao, Chengyue Wang, Kaiqun Zhu, Yilong Ren, Bolin Gao, Shengbo Eben Li, Chengzhong Xu, and Zhenning Li. 2025. Minds on the move: Decoding trajectory prediction in autonomous driving with cognitive insights. IEEE Transactions on Intelligent Transportation Systems (2025)

    2025

  32. [40]

    Parsa Mahmoudieh, Deepak Pathak, and Trevor Darrell. 2022. Zero-shot re- ward specification via grounded natural language. In International Conference on Machine Learning. PMLR, 14743–14752

    2022

  33. [41]

    Doaa Mahmud, Hadeel Hajmohamed, Shamma Almentheri, Shamma Alqaydi, Lameya Aldhaheri, Ruhul Amin Khalil, and Nasir Saeed. 2025. Integrating LLMs with ITS: Recent Advances, Potentials, Challenges, and Future Directions. IEEE Transactions on Intelligent Transportation Systems (2025)

    2025

  34. [42]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. 2024. Grounding dino: Marry- ing dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision . Springer, 38–55

    2024

  35. [43]

    Muhammad Monjurul Karim, Yu Li, and Ruwen Qin. 2021. Towards explainable artificial intelligence (XAI) for early anticipation of traffic accidents.arXiv e-prints (2021), arXiv–2108

    2021

  36. [44]

    Farzeen Munir, Shoaib Azam, Tsvetomila Mihaylova, Ville Kyrki, and Tomasz Pi- otr Kucner. 2025. Pedestrian Vision Language Model for Intentions Prediction. IEEE Open Journal of Intelligent Transportation Systems (2025)

    2025

  37. [45]

    Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. 3195–3204

    2019

  38. [46]

    World Health Organization. 2023. Global status report on road safety 2023: sum- mary. World Health Organization

    2023

  39. [47]

    Alexandros Papadopoulos, Athanasios Sersemis, Georgios Spanos, Antonios Lalas, Christos Liaskos, Konstantinos Votis, and Dimitrios Tzovaras. 2024. Light- weight accident detection model for autonomous fleets based on GPS data.Trans- portation research procedia 78 (2024), 16–23

    2024

  40. [48]

    National Highway Traffic Safety Administration. 2023. NHTSA Early Estimates: 2022 Traffic Crash Deaths. https://www.nhtsa.gov/press-releases/traffic-crash- death-estimates-2022. Accessed: 2025-03-04

    2023

  41. [49]

    Liang Shi, Boyu Jiang, Tong Zeng, and Feng Guo. 2025. ScVLM: Enhancing Vision-Language Model for Safety-Critical Event Understanding. In Proceedings of the Winter Conference on Applications of Computer Vision . 1061–1071

    2025

  42. [50]

    Xiao Teng, Zhenjiang Shen, Lin Huang, Hui Li, and Wankai Li. [n. d.]. Improving Intelligent Perception and Decision Optimization of Pedestrian Crossing Sce- narios in Autonomous Driving Environments Through Large Visual Language Models. A vailable at SSRN 5112947 ([n. d.])

  43. [51]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, I. Sutskever, T. Salimans, and D. Amodei. 2021. Learning Transferable Visual Models From Natural Language Supervision. Proceedings of the International Conference on Machine Learning (ICML) 139 (2021), 8748–8763...

    2021 arXiv

  44. [52]

    Sebastian Wandelt, Changhong Zheng, Shuang Wang, Yucheng Liu, and Xiaoqian Sun. 2024. Large language models for intelligent transportation: A review of the state of the art and challenges. Applied Sciences 14, 17 (2024), 7455

    2024

  45. [53]

    Tianhang Wang, Kai Chen, Guang Chen, Bin Li, Zhijun Li, Zhengfa Liu, and Changjun Jiang. 2023. GSC: A graph and spatio-temporal continuity based framework for accident anticipation. IEEE Transactions on Intelligent Vehicles 9, 1 (2023), 2249–2261

    2023

  46. [54]

    Ayush Verma and Manju Khari. 2024. Vision-Based Accident Anticipation and Detection Using Deep Learning. IEEE Instrumentation & Measurement Magazine 27, 3 (2024), 22–29

    2024

  47. [55]

    Chaowu Xie, Kun Zhang, Jiangchi Zhang, and Songshan Huang. 2025. Driving factors and spatio-temporal evolution of tourist accident clusters in Chinese outbound tourism. Current Issues in Tourism (2025), 1–19

    2025

  48. [56]

    Yu Yao, Mingze Xu, Yuchen Wang, David J Crandall, and Ella M Atkins. 2019. Unsupervised traffic accident detection in first-person videos. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 273–280

    2019

  49. [57]

    Kajsa Weibull, Björn Lidestam, and Erik Prytz. 2025. Driver’s gaze behavior when approached by an emergency vehicle–The effects of in-car warnings and system introduction. Transportation Research Part F: Traffic Psychology and Behaviour 109 (2025), 137–146

    2025

  50. [58]

    Jiaxun Zhang, Yanchen Guan, Chengyue Wang, Haicheng Liao, Guohui Zhang, and Zhenning Li. 2025. LATTE: Lightweight Attention-based Traffic Accident Anticipation Engine. arXiv:2504.04103 [cs.CE] https://arxiv.org/abs/2504.04103

    2025 arXiv

  51. [59]

    Yahui ZHANG, Ying LI, and Tianen LIU. 2025. Traffic Accident Prediction Method Based on Local Relational Features and Attention Mechanisms. Transactions of Beijing institute of Technology 45, 1 (2025), 11–18

    2025

  52. [60]

    Kuo-Hao Zeng, Shih-Han Chou, Fu-Hsiang Chan, Juan Carlos Niebles, and Min Sun. 2017. Agent-centric risk assessment: Accident anticipation and risky region localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2222–2230

    2017

  53. [63]

    Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. 2023. Tame a wild camera: in-the-wild monocular camera calibration. Advances in Neural Information Processing Systems 36 (2023), 45137–45149

    2023

  54. [2023]

    Accident Analysis & Prevention 180 (2023), 106923

    Automated vehicle data pipeline for accident reconstruction: New insights from LiDAR, camera, and radar data. Accident Analysis & Prevention 180 (2023), 106923

    2023

  55. [2025]

    Computers and Electronics in Agriculture 229 (2025), 109611

    Boosting grape bunch detection in RGB-D images using zero-shot annota- tion with Segment Anything and GroundingDINO. Computers and Electronics in Agriculture 229 (2025), 109611

    2025

Pith tools

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