AI summary · arXiv: 2611.18442v1 · cs.CR

CANTraj: Recovering Vehicle Trajectories from CAN Signals in Naturalistic Driving

Yuhang Wang, Jingxin Yang, Chuheng Wei, Suyang Chen, Haozhuang Chi, Hao Zhou
Original paper Submitted to arXiv on 13 Nov. 2026 License: NONEXCLUSIVE-DISTRIB 1.0 Also in: fr

How to read this summary? Illustrated guide

The essentials

CANTraj shows that vehicle CAN signals, especially steering combined with road maps, can recover naturalistic driving routes despite GPS removal.

Question

How accurately can existing methods reconstruct vehicle trajectories from CAN signals, and how do signal choice, map access, initial heading, and online versus offline processing affect recovery? The study also examines method failures and whether quantising signals reduces the demonstrated location leakage.

Method

The authors introduce CANTraj, containing 3,940 drives from 454 drivers, and evaluate CANTraj-Bench on 3,675 drives from 431 drivers, spanning 1,380 hours and 85,919 km. They re-implement six published attacks and compare them with map-free baselines, an online particle filter, offline beam search, and learned models under specified information protocols; the main evaluation includes 776 drives from 94 unseen drivers. Evaluation uses route recovery and position-error metrics, counts missing outputs as failures, and includes paired comparisons and driver-resampled confidence intervals.

Findings

  • Result On five-minute test windows with known start and heading, adding steering to the online map filter improves route recovery from 56% with speed alone to 87%, while median ADE falls from 286 m to 43 m.
    See the passage in the paper
    The untrained map filter gives a consistent result on all test windows: 43 m with steering versus 286 m with speed alone (paired difference $-$59.5 m).
  • Result Offline beam search with steering recovers 88% of five-minute test routes with the heading known and 81% when it is withheld.
    See the passage in the paper
    With steering, the start heading barely moves the offline beam’s median, unlike the filter’s; it still matters in the tail and, with speed alone, for the beam too.
  • Result Without a road map, the tested CAN-only methods drift substantially over five minutes; nominal steering dead reckoning has a 921 m median ADE on test windows.
    See the passage in the paper
    Nominal steering dead reckoning has a median ADE of 921 m on test and 1,669 m on few-shot drivers, no better than a straight line at the measured speed (735 m);
  • Authors’ interpretation Quantising steering to 10° or speed to 1 m/s leaves the tested map attacks largely effective, although the authors caution that these results concern fixed attacks rather than attackers adapted to degraded signals.
    See the passage in the paper
    Steering quantised to 10∘, offset by 5∘ or held at 1 Hz changes the route recovery of either attack by at most 1.3 percentage points;
  • Result Different reconstruction methods succeed on different windows: an oracle selecting among ten map methods recovers routes in 98% of test windows, compared with 88% for the beam.
    See the passage in the paper
    Some method recovers the route in 98 % of the test windows (few-shot 98 %) against 88 % for the beam, our fixed reference, a difference of 10 points (CI [6, 16]);

Limitations

  • The dataset and results are geographically and technically uneven: North American roads and vehicles supported by the logging system dominate, while some signals are available only for subsets.
    See the passage in the paper
    CANTraj broadens the range of drivers and vehicles available for CAN trajectory evaluation, but North American roads and vehicles supported by the logging system dominate.
  • The task primarily evaluates recovery from a known starting position, so the area-prior experiments do not establish performance for unrestricted unknown-start tracking.
    See the passage in the paper
    The task mainly measures recovery after one position is known.
  • The study tests fixed attacks on degraded signals, so it does not establish that quantisation provides protection against an attacker who adapts to those changes.
    See the passage in the paper
    These are sensitivities of fixed attacks, none re-tuned, re-calibrated or re-trained on the degraded signals: they bound what a publisher gains from coarsening against these attacks, not against an adapted one, and set no protection threshold.

Why it matters

The authors argue that removing GPS coordinates or initial heading does not eliminate route information in vehicle signals, and that steering combined with road-network constraints can enable substantial location recovery. CANTraj-Bench provides a shared evaluation framework for comparing attacks, signal contributions, and failure-aware recovery.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: Vehicle trajectory reconstruction; CAN bus signals; Route inference; GNSS privacy; Road-network matching

Read further

  • Comprehensive Summary
  • Layman's Summary
  • Blog article
  • Original abstract
The paper examines how accurately vehicle trajectories can be reconstructed from controller area network (CAN) signals when GPS coordinates are unavailable. It asks whether existing route-inference methods can be fairly compared on a broad, realistic dataset and how their accuracy varies with available signals, map access, initial heading, and observation-window length. To support this evaluation, the authors introduce CANTraj, a dataset pairing decoded CAN signals with time-aligned GNSS tracks from 3,940 drives by 454 drivers. CANTraj-Bench includes 3,675 drives by 431 drivers, spanning 1,380 hours and 85,919 kilometres. The main evaluation covers 776 drives by 94 drivers not represented in training. The benchmark compares six re-implemented published attacks with baselines for dead reckoning, map-based particle filtering, and offline beam search. Each method receives a defined set of inputs: a known starting position, sometimes an initial heading, selected CAN signals, and, in map-based protocols, a local road network. The study evaluates online methods, which process signals as they arrive, and offline methods, which can use the full observation window. It measures position error, the time spent within 100 metres of the reference, route recovery, and failures. Methods that return no result count as failures rather than being excluded. The authors also investigate different signal combinations, observation-window durations, degraded signal quality, and an LLM-based system that selects among candidate reconstructions. The largest gains come from combining steering data with road-network constraints. On five-minute test windows, an online map filter using speed alone has a median average displacement error of 286 metres and recovers 56% of routes. Adding steering reduces the median error to 43 metres and increases route recovery to 87%. Without a map, steering-based dead reckoning has a median error of 921 metres. Offline beam search using speed and steering recovers 88% of test routes, with a median error of 37 metres when the starting heading is known. Without that heading, its median error is 41 metres and route recovery falls to 81%. The online filter is more sensitive to missing heading: its median error rises from 43 to 138 metres. Offline access also helps avoid some large errors on longer windows, although the beam search and filter have similar median performance in several paired comparisons. Signal-combination experiments show that steering provides substantial value, while continuous accelerator-pedal data and pedal-pressed flags generally add little when speed is already available. Learned motion models improve map-free tracking in some settings but do not improve the map-based filter. Results for earlier attacks vary: OPD-II achieves a median error of 42 metres on test windows, whereas SPy performs poorly under the benchmark’s adaptations. CAN-Trace produces few matches, and many routes in its area-based adaptation fall outside the permitted search area. These findings apply to the specific implementations and protocols tested; they do not demonstrate that the original methods fail in every setting. The authors conclude that withholding GPS or initial heading does not eliminate location leakage from vehicle signals. Quantising steering to 10 degrees or speed to 1 metre per second leaves the tested map-based attacks largely effective. However, CANTraj is dominated by North American roads and supported vehicles, and results from a small controlled supplement show that method rankings can vary across settings. Some signals are available only for subsets of vehicles; the evaluation generally assumes a known starting position; and several method adaptations differ from their original tasks. The authors emphasise that failures by the tested attacks do not guarantee privacy and that protection measures should be assessed against attackers adapted to the modified data.

Was this synthesis useful?

Score: 0

Why do we need votes?

Votes are used to determine whether we need to re-run our summarizing tools. If the count reaches -10, our tools can be restarted.

Figures and plots

Ask about a figure by number or use its question button to request visual analysis. Up to two figures and four image panels per question. Values read from plots are approximate; check the original data.

AI assistant instructions?

Tools for this paper

Explain it differently 1 free rewrite left today
Save to a collection Group papers in collections to compare them, chat with them and export them. Continue with Google Create a free account Log in

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero Zotero guide
Citation alerts Citation alerts guide
Weekly letter Presented in "Computing beyond AI this week" of 16 November 2026. Computing beyond AI this week: The costs hidden in our maps, networks and algorithms · Receive one by email for free

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

  1. Close Broader context
    CAN Coach: Vehicular Control through Human Cyber-Physical Systems
    M. Nice, S. Elmadani, R. Bhadani et al. · 2021 · Robotics · cited 16 times · AI summary available

    CAN Coach uses CAN radar signals to provide driver feedback for longitudinal control, whereas CANTraj uses CAN signals to reconstruct vehicle routes.

    Abstract

    This work addresses whether a human-in-the-loop cyber-physical system (HCPS) can be effective in improving the longitudinal control of an individual vehicle in a traffic flow. We introduce the CAN Coach, which is a system that gives feedback to the human-in-the-loop using radar data (relative speed and position information to objects ahead) that is available on the controller area network (CAN). Using a cohort of six human subjects driving an instrumented vehicle, we compare the ability of the human-in-the-loop driver to achieve a constant time-gap control policy using only human-based visual perception to the car ahead, and by augmenting human perception with audible feedback from CAN senso…

  2. Close Broader context
    Improving Autonomous Vehicle Mapping and Navigation in Work Zones Using Crowdsourcing Vehicle Trajectories
    Hanlin Chen, Renyuan Luo, Yiheng Feng · 2023 · Robotics · preprint, not peer reviewed · cited 6 times · AI summary available

    This work uses crowdsourced vehicle trajectories to improve mapping and navigation in work zones, while CANTraj studies recovering trajectories from CAN signals.

    Abstract

    Prevalent solutions for Connected and Autonomous vehicle (CAV) mapping include high definition map (HD map) or real-time Simultaneous Localization and Mapping (SLAM). Both methods only rely on vehicle itself (onboard sensors or embedded maps) and can not adapt well to temporarily changed drivable areas such as work zones. Navigating CAVs in such areas heavily relies on how the vehicle defines drivable areas based on perception information. Difficulties in improving perception accuracy and ensuring the correct interpretation of perception results are challenging to the vehicle in these situations. This paper presents a prototype that introduces crowdsourcing trajectories information into the …

  3. Close Broader context
    Road Genome: A Topology Reasoning Benchmark for Scene Understanding in Autonomous Driving
    Huijie Wang, Zhenbo Liu, Yang Li et al. · 2023 · Computer Vision and Pattern Recognition · cited 127 times · AI summary available

    Road Genome benchmarks topology reasoning for autonomous-driving scene understanding, while CANTraj benchmarks recovery of vehicle trajectories from CAN signals.

    Abstract

    Understanding the complex traffic environment is crucial for self-driving vehicles. Existing benchmarks in autonomous driving mainly cast scene understanding as perception problems, e.g., perceiving lanelines with vanilla detection or segmentation methods. As such, we argue that the perception pipeline provides limited information for autonomous vehicles to drive in the right way, especially without the aid of high-definition (HD) map. For instance, following the wrong traffic signal at a complicated crossroad would lead to a catastrophic incident. By introducing Road Genome (OpenLane-V2), we intend to shift the community's attention and take a step further beyond perception - to the task of…

  4. Related Broader context
    NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking
    Daniel Dauner, Marcel Hallgarten, Tianyu Li et al. · 2024 · Computer Vision and Pattern Recognition · cited 549 times · AI summary available

    NAVSIM benchmarks vision-based driving policies using data and non-reactive simulation, whereas CANTraj benchmarks trajectory recovery from CAN signals.

    Abstract

    Benchmarking vision-based driving policies is challenging. On one hand, open-loop evaluation with real data is easy, but these results do not reflect closed-loop performance. On the other, closed-loop evaluation is possible in simulation, but is hard to scale due to its significant computational demands. Further, the simulators available today exhibit a large domain gap to real data. This has resulted in an inability to draw clear conclusions from the rapidly growing body of research on end-to-end autonomous driving. In this paper, we present NAVSIM, a middle ground between these evaluation paradigms, where we use large datasets in combination with a non-reactive simulator to enable large-sc…

  5. Related Broader context
    How Does It Feel? Self-Supervised Costmap Learning for Off-Road Vehicle Traversability
    Mateo Guaman Castro, Samuel Triest, Wenshan Wang et al. · 2022 · Robotics · cited 118 times · AI summary available

    This method learns off-road traversability costmaps from environmental data and terrain feedback, rather than reconstructing routes from CAN signals.

    Abstract

    Estimating terrain traversability in off-road environments requires reasoning about complex interaction dynamics between the robot and these terrains. However, it is challenging to build an accurate physics model, or create informative labels to learn a model in a supervised manner, for these interactions. We propose a method that learns to predict traversability costmaps by combining exteroceptive environmental information with proprioceptive terrain interaction feedback in a self-supervised manner. Additionally, we propose a novel way of incorporating robot velocity in the costmap prediction pipeline. We validate our method in multiple short and large-scale navigation tasks on a large, aut…

  6. Related Shared data, tool or method
    nuScenes: A multimodal dataset for autonomous driving
    Holger Caesar, Varun Bankiti, Alex H. Lang et al. · 2019 · Machine Learning · cited 9,545 times · AI summary available

    nuScenes provides multimodal autonomous-driving sensor data for training and evaluation, while CANTraj provides paired CAN signals and GNSS trajectories for recovery benchmarks.

    Abstract

    Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first dataset to carry the full autonomous vehicle sensor suite: 6 camera…

  7. Related Broader context
    Safe Occlusion-aware Autonomous Driving via Game-Theoretic Active Perception
    Zixu Zhang, Jaime F. Fisac · 2021 · Robotics · cited 95 times · AI summary available

    This work plans safe autonomous-vehicle trajectories under occlusion, while CANTraj reconstructs trajectories from vehicle signals.

    Abstract

    Autonomous vehicles interacting with other traffic participants heavily rely on the perception and prediction of other agents' behaviors to plan safe trajectories. However, as occlusions limit the vehicle's perception ability, reasoning about potential hazards beyond the field-of-view is one of the most challenging issues in developing autonomous driving systems. This paper introduces a novel analytical approach that poses the problem of safe trajectory planning under occlusions as a hybrid zero-sum dynamic game between the autonomous vehicle (evader), and an initially hidden traffic participant (pursuer). Due to occlusions, the pursuer's state is initially unknown to the evader and may late…

  8. Related Broader context
    Hidden Biases of End-to-End Driving Models
    Bernhard Jaeger, Kashyap Chitta, Andreas Geiger · 2023 · Computer Vision and Pattern Recognition · cited 175 times · AI summary available

    This work analyzes biases in end-to-end driving models and proposes alternatives, whereas CANTraj evaluates methods for reconstructing routes from CAN signals.

    Abstract

    End-to-end driving systems have recently made rapid progress, in particular on CARLA. Independent of their major contribution, they introduce changes to minor system components. Consequently, the source of improvements is unclear. We identify two biases that recur in nearly all state-of-the-art methods and are critical for the observed progress on CARLA: (1) lateral recovery via a strong inductive bias towards target point following, and (2) longitudinal averaging of multimodal waypoint predictions for slowing down. We investigate the drawbacks of these biases and identify principled alternatives. By incorporating our insights, we develop TF++, a simple end-to-end method that ranks first on …

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 16 November 2026.

The illustrated guide to the papers around a paper

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.