VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes

arXiv:2606.06074 · cs.CV · Submitted 2026-06-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes".

Title and authors: Tom: We've been looking at the VZCrash dataset, and it really highlights how much data volume impacts model generalization. Now let's look closer at what the paper actually introduces.

Jane: The paper is titled "VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes," and it sets itself up as a massive resource for collision detection in driving scenarios using Inertial Measurement Unit data.

Lu: To put it simply, they’ve created this dataset that contains over thirty thousand verified crashes and nearly one hundred fifty thousand negative samples designed to challenge any system trying to detect these events.

Meng: So, when we talk about the data itself, what exactly are we getting in terms of hardware and recording frequency? Is this something I can realistically ingest into a production pipeline?

Tom: Each sample in VZCrash includes sixteen seconds of data captured at one hundred Hertz for both tri-axial accelerometer and gyroscope readings. They also include one reading per second for GPS speed.

Jane: So, that level of detail—that high frequency measurement—is what gives the system the rich information needed to distinguish a true crash from just a sudden jolt on the road.

Lu: And they added some context to how this data was gathered; it came from a fleet of seventy-three thousand commercial vehicles driving in the United States over several years. This real-world capture is what makes it so valuable for testing.

Meng: seventy-three thousand vehicles? That’s a huge source. Does the paper mention how they handled the labeling process for those events?

Tom: They describe a specific process where events were collected on-device based on multiple threshold heuristics applied to acceleration signals: high values on the X axis are considered hard breaking and hard accelerations events, while high values in the XY norm are considered shock events.

Jane: So, it’s not just one single trigger; they use a set of rules to flag potential incidents, which then get processed further using an ensemble of specialized machine learning models on the cloud.

Lu: And those flagged events were then reviewed by expert human reviewers along with a random sample of the negative events to ensure quality control.

The paper's summary: Tom: So, what's the main takeaway from this study regarding how this dataset helps us? It boils down to scaling data volume is especially important for training high-quality crash detection models.

Jane: Exactly, Tom; it shows that the ability of models to generalize across diverse real-world conditions scales significantly with data volume. This means more examples mean better performance in messy situations.

Lu: The experiment demonstrates that when you increase the supervised training set size, the model robustness improves substantially because they show a clear improvement once you include one percent of the available supervised samples.

Meng: That’s good for our training pipelines, but what about those models that are already pretty good? Does adding more data help them more than it helps beginners?

Tom: The gap between the baseline and evaluated deep learning architectures is significant, so even a small increase in data helps bridge that gap considerably.

Jane: It suggests that for real-world deployment, you can’t just rely on a small set of training examples; you need the full picture to capture all the variance.

Lu: The paper also shows that performance in those real-world populations drops significantly when models are trained on smaller subsets, particularly for the Chronos-Bolt and CNN-RNN architectures.

Meng: So, if I'm deploying something now, should I worry about using less data than what we can afford?

Tom: You have to be careful; if you train a model on too little data, you risk losing its ability to handle the variety of real-world crashes that aren't perfectly represented.

Jane: It really stresses that for those specific architectures, the dependency on having enough examples is quite high when we want them to perform reliably outside of a controlled lab setting.

Lalam: I can confirm what the literature says: if you train on less data, the performance suffers noticeably in those cases because they don't learn enough patterns.

The paper's improvements: Tom: Now let’s talk about what the authors suggest as improvements for this whole VZCrash setup to make it even more effective. They are looking at ways to enhance detection accuracy by incorporating other sensor data.

Jane: What kind of extra data are they thinking about adding? Are we talking more from the IMU, or maybe something external like GPS?

Lu: They suggest that integrating gyroscope and GPS speed data could enhance detection accuracy, particularly in those complex edge cases where the accelerometer data alone might be ambiguous.

Meng: That sounds promising for robustness; if the acceleration signal is noisy, having supplementary speed data could help confirm what’s happening. What does that mean for deployment complexity?

Tom: It means we're looking at a way to make the detection process more reliable when things get weird in the field. It adds another layer of confirmation to the decision-making process.

Jane: Beyond just adding data, what about improving how we label these events themselves? Are they thinking about moving beyond a simple yes or no answer?

Lu: Yes, future updates to VZCrash will focus on increasing label granularity, shifting from binary crash detection to labeling specific event types, such as the direction of impact or the severity of the collision.

Meng: That sounds like a huge undertaking for data annotation teams. It’s not just about counting crashes anymore; it's about detailed scene understanding.

Tom: It changes how we view this data; it’s not just a binary flag anymore; it’s becoming more descriptive information about the event itself.

Jane: And this aligns perfectly with what we need to solve the problem of noise and extreme class imbalance inherent in real-world telemetry that these models face.

Lu: Ultimately, the goal is to release VZCrash to help develop AI systems that are better equipped to handle those difficult real-world scenarios without needing perfect data from the start.

Conclusion: Tom: So, wrapping up on "VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes," the main message is that massive data volume improves model generalization, especially for deep learning architectures.

Jane: That means we see a significant performance gain when scaling training sets to use the full dataset because real-world conditions have a wide range of variance.

Lu: The paper shows that relying on smaller subsets causes performance to drop noticeably in specific models like Chronos-Bolt and CNN-RNN if you want them to stay reliable.

Meng: For me, the practical conclusion is that we need large datasets for any serious deployment, even if it means investing in getting more data or better labeling.

Lalam: I'm confident that this dataset will be a critical resource for pushing the development of models capable of handling the noise and imbalance in real-world telemetry.

Tom: It’s a big step forward because this work shows that scale really matters when you want reliable results from these types of data.

Jane: It’s an important piece of infrastructure if we want to move toward safer autonomous driving systems that can handle the complexity we see on the road today.

Verizon Connect

cs.CV

Submitted: 2026-06-04

Updated: 2026-11-17

Comments: Accepted at the 2026 IEEE International Conference on Intelligent Transportation Systems (ITSC 2026). Revised after the conference to fix typos and inconsistencies. VZCrash is publicly available at this URL: https://huggingface.co/datasets/vzc-research-chapter/VZCrash

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: VZCrash introduces VZCrash, the largest publicly available dataset of real-world vehicle collisions featuring highfrequency IMU telemetry The dataset contains more than 31,000 validated crashes and…

Key concepts

VZCrash Dataset
This is a massive dataset containing nearly 190,000 samples of vehicle events captured from commercial vehicles across the US. Each event includes 16 seconds of high-frequency (100 Hz) accelerometer and gyroscope data, plus GPS speed information. It contains over 31,090 verified crashes.
Physical Baseline
This is a learning-free method that detects a crash by measuring the peak across time of the vector norm of acceleration along the vehicle's longitudinal and transversal axes. It serves as a simple, fast benchmark against which more complex deep learning models are compared to assess their performance.
Data Scaling Impact
The study demonstrates that increasing supervised training data volume substantially improves model robustness and generalization. Performance gaps between baseline methods and deep learning architectures become more significant when models are trained on smaller subsets of the available samples.

Terminology

Summary

VZCrash introduces VZCrash, the largest publicly available dataset of real-world vehicle collisions featuring highfrequency IMU telemetry The dataset contains more than 31,000 validated crashes and 158,000 negative samples This work presents an extensive experimental study enabled by the volume of the dataset The results demonstrate that the ability of models to generalize across diverse real-world conditions scales significantly with data volume. p. 0

Dataset Composition and Labeling

VZCrash comprises almost 190,000 samples, including 31,090 verified crashes For each event, 16 seconds of 100 Hz tri-axial accelerometer and gyroscope data is released along with 1 Hz GPS speeds. Events were captured by devices installed on a fleet of 73,010 commercial vehicles across the United States over several years. p. 0 A crash is defined as an event involving the ego-vehicle that satisfies at least one of the following criteria: Unintentional contact or accidental departure They distinguish crashes from other high-intensity acceleration events exhibiting similar kinematic profiles, such as pothole encounters or commercial trailer docking procedures. p. 1

Model Architectures Benchmarked

The experimental study benchmarks several different approaches on the crash detection task The physical baseline is a learning-free approach defined by the peak across time of the vector norm of acceleration along the longitudinal and transversal vehicle axes. Other models include CNN-RNN, 1D Swin Transformer, CNN-Transformer, Chronos-Bolt, and Scalogram Classifier. p. 4 The CNN-Transformer architecture is noted as the most efficient among the evaluated models by a factor of 3× compared to others. p. 3

Impact of Data Scaling

The results show that increasing the supervised training set to the full dataset leads to substantial improvements in model robustness. The performance gap between the baseline and evaluated deep learning architectures is significant When evaluating model sensitivity to data volume, AP scores improve as the training set size increases from 1% up to 100% of the available supervised samples. p. 3 However, performance in real-world event populations drops significantly when models are trained on smaller subsets, especially for Chronos-Bolt and CNN-RNN. p. 4

Real-World Deployment Performance

When tested under realistic deployment conditions with a long-tail distribution of events, the Average Precision of the best model drops from 97.5% on the VZCrash test set to around 86% in the wild. p. 4 This experiment underscores that a model must maintain near-perfect precision to avoid detecting a huge amount of false positives in a stream where non-crash events outnumber actual collisions by a factor of 5,100 to 1. p. 5 The physical baseline achieves an AP of nearly 90% on the VZCrash test set. p. 3 The marginal utility of new examples decreases as one gets closer to the full size, but both CNN-RNN and CNN-Transformer have still not fully plateaued. p. 4

Future Directions

Integrating gyroscope and GPS speed data could enhance detection accuracy, particularly in complex edge cases where accelerometer data alone may be ambiguous. Future updates to VZCrash will focus on increasing label granularity, moving from binary crash detection to labeling specific event types such as the direction of impact or the severity of the collision. p. 5 The release of VZCrash is aimed at facilitating the development of models better equipped to handle noise and extreme class imbalance inherent in real-world telemetry. p. 5

References

[1] J. M. Hankey, M. A. Perez, and J. p. 5 A. McClafferty, “Description of the SHRP 2 naturalistic database and the crash, near-crash, and baseline data sets,” Virginia Tech Transportation Institute, Tech. p. 5 Rep., 2016

[2] H. Caesar, V. Bankiti, A. p. 5 H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y Pan, G Baldan, and O Beijbom “nuScenes: A Multimodal Dataset for Autonomous Driving,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Seattle WA USA: IEEE Jun 2020 pp 11 618–11 628

[3] F Yu, H Chen, X Wang, W Xian, Y Chen, F Liu, V Madhavan, and T Darrell “BDD100k: A diverse driving dataset for heterogeneous multitask learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2020 pp 2636–2645

[4] D Moura, S Zhu, and O Zvitia “Nexar dashcam collision prediction dataset and challenge,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops 2025 pp 2583–2591

[5] J Fang, L-L Li, K Yang, Z Zheng, J Xue, and T-S Chua “Cognitive accident prediction in driving scenes: A multimodality benchmark” arXiv preprint arXiv:2212.09381 2022

[6] S Moosavi, M H Samavatian, S Parthasarathy, R Teodorescu, and R Ramnath “Accident risk prediction based on heterogeneous sparse data: New dataset and insights” in Proceedings of the 27th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems 2019 pp 33–42

[7] National Highway Traffic Safety Administration “Fatality analysis reporting system (FARS)” https://www.nhtsa.gov/research-data/fatality-analysis-reporting-system-fars 2023 accessed 2026-02-24

[8] L Kubin, T Bianconcini, D C de Andrade, M Simoncini, L Taccari, and F Sambo “Deep crash detection from vehicular sensor data with multimodal self-supervision” IEEE Transactions on Intelligent Transportation Systems vol 23 no 8 pp 12 480–12 489 2021

[9] L Taccari, F Sambo, L Bravi, S Salti, L Sarti, M Simoncini, and A Lori “Classification of crash and near-crash events from dashcam videos and telematics” in 2018 21st International Conference on intelligent transportation systems (ITSC) IEEE 2018 pp 2460–2465

[10] K Cho, B Van Merrienboer, C Gulcehre, D Bahdanau, F Bougares, H Schwenk, and Y Bengio “Learning phrase representations using rnn encoder–decoder for statistical machine translation” in Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) 2014 pp 1724–1734

[11] L Shi, Y Chen, M Liu, and F Guo “Dust: Dual swin transformer for multi-modal video and time-series modeling” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2024 pp 4537–4546

[12] Z Liu, Y Lin, Y Cao, H Hu, Y Wei, Z Zhang, S Lin, and B Guo “Swin transformer: Hierarchical vision transformer using shifted windows” in Proceedings of the IEEE/CVF International Conference on Computer Vision 2021 pp 10 012–10 022

[13] A F Ansari, L Stella, C Turkmen, X Zhang, P Mercado, H Shen, O Shchur, S Rangapuram, S Arango, S Kapoor et al “Chronos: Learning the language of time series” arXiv preprint arXiv:2403.07815 2024

[14] Q Kong, Y Cao, T Iqbal, Y Wang, W Wang, and M D Plumbley “Panns: Large-scale pretrained audio neural networks for audio pattern recognition” IEEE/ACM Transactions on Audio Speech and Language Processing vol 28 pp 2880–2894 2020

[15] Y Gong, Y A Chung, and J Glass “Ast: Audio spectrogram transformer” arXiv preprint arXiv:2104.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve crash detection robustness by training models on larger datasets, as the ability of models to generalize across diverse real-world conditions scales significantly with data volume. This allows deep learning models to better capture the variance of real-world crashes, including difficult to observe edge cases.

  2. Enhance low-latency deployment capability by leveraging the Physical Baseline approach, which is described as a learning-free approach inspired by what is commonly done both in GPS tracking devices to detect 'impulses,' and algorithms that control airbag deployment. This provides a fast method for real-time detection suitable for power-constrained environments.

  3. Increase accuracy in real-world deployment scenarios by employing an ensemble-informed labeling strategy, where we review the 1,000 events that received the highest confidence scores across all models previously trained on VZCrash. This helps identify actual crashes from the harsh driving events (HDEs) in a way that maintains high precision when actual collisions are extremely rare.

  4. Improve model generalization under data scarcity by using smaller subsets of the training data and testing on the full test set, as a clear improvement appears once we include at 0.2k (1%) [of training samples]. This determines the minimum data threshold required to distinguish true collisions from kinematic distractors for high-capacity architectures.

  5. Enhance detection in noisy environments by utilizing the Scalogram Classifier with MobileNetV3 Small, as it is inspired by audio pattern recognition and uses a time-frequency representation of the signal obtained by taking the magnitude of their continuous wavelet transform (CWT). This allows classification even when raw acceleration data contains high-frequency vibrations or noise.

Sources

Related papers