exiD: Drone-Based Highway Trajectories
- exiD is a drone-based naturalistic trajectory dataset capturing vehicle movements on highway ramps with rich merging, diverging, and weaving behaviors.
- It comprises approximately 69,000 trajectories from multiple Autobahn sites with detailed kinematics and map context in OpenDRIVE and Lanelet2 formats.
- The dataset supports benchmarking for trajectory forecasting, lane-change intention prediction, adaptive decision-making, and robust motion planning in interactive highway scenarios.
exiD is a drone-based naturalistic trajectory dataset collected over highway on-ramp and off-ramp areas in Germany and repeatedly used as a benchmark for highly interactive highway scenarios, especially merging, diverging, and weaving. Later studies describe it as following the same aerial recording philosophy as highD while providing richer geometry, denser interactions, and ramp-related maneuvers, which makes it useful not only for trajectory forecasting but also for motion planning, lane-change intention prediction, adaptive decision-making, and traffic-flow analysis (Mozaffari et al., 2023, Shi et al., 30 Dec 2025, Shi et al., 22 Sep 2025).
1. Dataset identity and distinguishing characteristics
In later literature, exiD is described as a drone-based naturalistic trajectory dataset collected over highway on-ramp and off-ramp areas, and as a real, accurate drone-based trajectory dataset recorded at German highway entries and exits. One study reports that it contains around 69,000 trajectories from seven Autobahn sites and includes detailed vehicle kinematics, metadata, and map context in OpenDRIVE and Lanelet2 formats; another reports 92 aerial video sequences, about 16 hours, about 69,172 vehicles, and trajectories processed at 25 Hz (Shi et al., 30 Dec 2025, Shi et al., 22 Sep 2025).
A recurring theme in the literature is that exiD is not treated as merely another highway corpus. Multiple papers explicitly contrast it with highD: highD is used as a straight-road or standard highway benchmark, whereas exiD is used as the ramp-rich, interaction-heavy benchmark in which merging and diverging are central to the task definition (Mozaffari et al., 2023, Shi et al., 22 Sep 2025). This distinction is operational rather than rhetorical: exiD is repeatedly selected when the target problem depends on vehicles entering or leaving the main carriageway, on nontrivial lane connectivity, or on dense local interactions near entries and exits (Mozaffari et al., 2023, Huang et al., 2024).
The dataset is also described as heterogeneous. One traffic-flow study states that exiD consists of four vehicle types—truck, van, car, and motorcycle—and categorizes it as external-view, highway, uninterrupted flow, camera-based, with no AVs involved (Leungbootnak et al., 5 Nov 2025). Other studies emphasize the opposite end of the same spectrum: not uninterrupted-flow abstraction, but highly interactive ramp behavior with merging, diverging, and weaving (Shi et al., 30 Dec 2025, Shi et al., 22 Sep 2025). This suggests that exiD supports both aggregate traffic analysis and maneuver-level interaction analysis, depending on the modeling objective.
2. Roadway geometry, semantics, and scenario structure
The scenarios most often associated with exiD are highway entries and exits with ramp merging/diverging zones. In one forecasting-and-planning study, exiD is characterized as containing large-scale, naturalistic, drone-recorded trajectories at highway entries and exits in Germany, with multiple locations, dense traffic, and merging geometry involving slip roads and multi-lane carriageways (Mozaffari et al., 2023). In a lane-change intention study, exiD is described as focusing on complex merging, diverging, and weaving behavior, with richer and more ambiguous interactions, more heterogeneous geometry, and ramp-related maneuvers whose semantics cannot be inferred from lane IDs alone (Shi et al., 30 Dec 2025).
This geometric specificity directly affects annotation logic. Several papers note that lane IDs in exiD are not sequential in the same way as on straight highways, especially across ramps and mainline lanes. As a result, left/right lane-change direction is often derived from lateral velocity rather than lane-ID transitions. One study defines the start time of a lane change as the moment when the lateral position crosses the current lane centerline by at least $0.2$ m and is followed by monotonic lateral drift for at least $0.5$ s, and the end time as the moment when the vehicle is fully within the adjacent lane with no reverse lateral motion for at least $1.0$ s afterward; for exiD specifically, the mean latVelocity over a $0.1$ s window starting at the lane-change start time determines whether the maneuver is labeled Left-LC or Right-LC (Shi et al., 30 Dec 2025). A closely related three-class framework uses the same exiD-specific lateral-velocity rule because mainline lanes and ramps do not have consistent sequential indexing (Shi et al., 22 Sep 2025).
Map semantics are also unusually salient in exiD-based work. Later studies explicitly derive ramp-related descriptors from OpenDRIVE annotations, including distance to ramp entry/exit, estimated time-to-arrival, and multi-horizon indicators at 5 s, 15 s, and 30 s (Shi et al., 30 Dec 2025, Shi et al., 22 Sep 2025). In forecasting studies, lane semantics are encoded with categories such as no lane, normal lane, expected merging lane, and merge lane, and these representations are reported as particularly informative because lane semantics and merge topology strongly shape future motion (Mozaffari et al., 2023).
3. Common preprocessing and experimental protocols
A notable feature of exiD research is that preprocessing is usually task-specific and geometry-aware rather than generic. For trajectory prediction in merging, one study converts trajectories to a Frenet frame, defines separate reference paths for the slip road and the main carriageway, takes the intersection point of the two reference paths as the origin, and projects each vehicle state into along-track and cross-track coordinates . The same study downsamples the original 25 Hz data to 5 Hz for prediction and uses variable-length observations from 2 to 15 time steps, corresponding to 0.4 s to 3 s of history, with a prediction horizon of 25 frames, i.e. 5 s (Mozaffari et al., 2023). A multimodal transformer study also uses exiD through single-lane merging scenarios from four locations and converts trajectories and lane markings from Cartesian to Frenet coordinates (Mozaffari et al., 2023).
Location-based evaluation is common when exiD is used to test generalization. A physics-informed lane-change intention paper trains on locations 0, 1, 2, and 3 and tests on locations 4, 5, and 6, with Bayesian hyperparameter search and 5-fold cross-validation inside the training portion (Shi et al., 30 Dec 2025). A distribution-shift study likewise exploits exiD’s seven locations with no geographic overlap by placing data from six maps in the ID split and preserving the last map for the OOD split; one sample constitutes 4 s of future with 2 s of history sampled at $\Delta t=\SI{0.2}{\second}$ (Diehl et al., 2024).
Other protocols are more localized. In the motion-planning study built around exiD merging, only the merging data of four locations are used; those locations all feature a single-lane slip road merging into a two- or three-lane main road. From these four locations, 49 recordings are selected in total, with 45 recordings for training and 4 recordings for evaluation of the prediction model, one test recording per location. Planning evaluation is then constructed from recording number 39 and built from 97 merging vehicles (Mozaffari et al., 2023). By contrast, a recurrent lane-change forecasting study on ramp areas uses an 80:20 train:test split, a 5-frame observation history, eight surrounding cars, and balanced classes, but does not specify a stronger scene-wise or trajectory-wise split protocol in the provided description (Abouras et al., 21 Jan 2026).
| Study | exiD task | exiD-specific setup |
|---|---|---|
| (Mozaffari et al., 2023) | Trajectory prediction and merge planning | Four merging locations; 45 train and 4 evaluation recordings; Frenet frame; 0.4–3 s observation; 5 s prediction |
| (Shi et al., 30 Dec 2025) | Three-class lane-change intention prediction | Train on locations 0–3, test on 4–6; left/right by averaged latVelocity; history windows 2–7 s |
| (Diehl et al., 2024) | Domain adaptation for structured multi-agent policy | Six maps ID, one map OOD; 2 s history, 4 s future, s |
These protocols show that exiD is rarely used as a raw trajectory store. Instead, its ramp geometry, lane semantics, and multi-location structure are usually elevated into the experimental design.
4. Benchmark roles in forecasting and intention prediction
In trajectory forecasting, exiD is often the dataset on which interaction-rich, merge-specific behavior is tested after model competitiveness is established elsewhere. In the variable-observation-length transformer study, highD serves primarily as the standard benchmark against published baselines, whereas exiD is the central dataset for assessing prediction across observation lengths and for studying downstream motion planning and control in merging (Mozaffari et al., 2023). On exiD, the reported RMSE values for constant velocity are 0.25 m, 0.63 m, 1.19 m, 1.92 m, and 2.82 m at 1 s through 5 s, while the proposed POVL model reports 0.17 m, 0.41 m, 0.76 m, 1.21 m, and 1.75 m. The same study states that even with only 0.4 s of observation, i.e. 2 time steps, the proposed method outperforms constant velocity (Mozaffari et al., 2023).
A multimodal transformer framework uses exiD specifically as a challenging merging benchmark that had not yet been exploited in multimodal prediction studies. For exiD, it reports minRMSE- values for MMnTP of 0.26, 0.57, 0.98, 1.50, and 2.11 at , and 0.17, 0.39, 0.69, 1.07, and 1.52 at , over 1 s through 5 s. The same paper reports MaxACC-$0.5$0 on exiD of 79.91, 89.54, 93.51, 95.08, 95.70, and 95.98 for $0.5$1 through $0.5$2, respectively (Mozaffari et al., 2023). Its exiD ablation further shows that a manoeuvre-based multimodal training loss preserves similar trajectory accuracy while reducing CollisionRate from 9.63% to 2.46% and OffroadRate from 26.83% to 0.06% relative to a standard multimodal trajectory loss (Mozaffari et al., 2023).
In lane-change intention prediction, exiD serves as the harder, more realistic stress test for three-class prediction. A temporal physics-informed multimodal framework reports exiD macro-F1 values of 0.9247, 0.8197, and 0.7605 at prediction horizons $0.5$3, 2, and 3 s, respectively, outperforming standalone LightGBM and Bi-LSTM baselines in that ramp-rich setting (Shi et al., 30 Dec 2025). A related physics-informed three-class framework reports exiD LightGBM accuracy and macro-F1 of 0.9607 and 0.9004 at 1 s, 0.8950 and 0.8343 at 2 s, and 0.8117 and 0.7819 at 3 s, with lower performance than on highD but a clear advantage over a two-layer stacked LSTM baseline (Shi et al., 22 Sep 2025). Both papers stress that exiD benefits from longer temporal context than highD, with the best LightGBM history window reported as $0.5$4 s for exiD (Shi et al., 30 Dec 2025, Shi et al., 22 Sep 2025).
A separate recurrent study isolates ramp-adjacent regions as an Area of Interest represented by ExiD and frames the task as three-class lane-change forecasting with labels LCL, LK, and LCR. Using a stacked two-layer LSTM and a 5-frame observation window, it reports ExiD AoI end-to-end accuracy, precision, recall, and F1 of 0.8279, 0.8204, 0.8289, and 0.8244 at 1 s, declining to 0.7688, 0.7534, 0.7287, and 0.7376 at 4 s (Abouras et al., 21 Jan 2026). The same paper highlights the gap between exiD-style ramp zones and ordinary highway segments by comparing about 76.9% accuracy in ExiD AoI against about 94.6% on normal highway sections at the 4-second horizon (Abouras et al., 21 Jan 2026).
5. Downstream use in planning, decision-making, and adaptation
One of exiD’s most consequential roles in the literature is as a bridge from prediction accuracy to motion-planning consequences. In the merge-planning study, the ego vehicle attempts to merge from a single-lane slip road onto the main carriageway, trajectories of nearby exiD vehicles are predicted one-by-one under an independence assumption, and those predicted futures are encoded into a potential field used by an MPC-based planner/controller. The planning horizon is 5 s, each scenario is initialized from a real merge trajectory, and three prediction inputs are compared: ground-truth future trajectories, constant velocity, and POVL (Mozaffari et al., 2023). Across all cases up to 50 m, replacing constant velocity with POVL yields 9.15% lower $0.5$5 score, 1.02% lower jerk, and 4.80% lower force according to the text, although the same paper notes that the table’s farthest-distance row shows 1.53% for the $0.5$6 m set (Mozaffari et al., 2023). In the densest bin, where surrounding vehicles are within 3 m of the ego trajectory, the reported relative improvements of POVL over CV are 19.18% in safety, 19.78% in comfort, and 13.20% in efficiency (Mozaffari et al., 2023).
exiD is also used to calibrate and test adaptive decision-making in ramp-merging interactions. A learning-enhanced game-theoretic approach uses highD and exiD to extract typical ramp-merging scenarios, calibrate interaction start and end moments, label yielding versus non-yielding behavior, learn reward-function parameters with maximum entropy IRL, and train a Bayesian-network mapping from environment variables to reward weights (Huang et al., 2024). After preprocessing, the combined corpus contains 1738 interactive vehicle data sequences, of which 1550 are used for training the mapping model and 188 for testing; the headline naturalistic-dataset result is 81.73% human-like decision similarity over 11,583 data points, and 77.12% similarity over 8,964 data points in 145 dynamic interactions (Huang et al., 2024). The same study does not report an exiD-only similarity score, which is an important limitation for exiD-specific interpretation (Huang et al., 2024).
In domain adaptation, exiD serves as a controlled geographical-shift benchmark for structured multi-agent policies. A low-rank residual decoder study uses exiD for open-loop joint prediction with seven locations, six maps for ID and one held-out map for OOD, and reports exact split sizes of 110,431/12,425/13,884 samples for ID train/val/test and 12,378/3,459/3,561 for OOD fine-tuning train/val/test (Diehl et al., 2024). On exiD, the strongest reported configuration is Fine-tuning + LoRD with Mix data, which yields OOD minSADE/minSFDE/bminSFDE/SMR of 0.86/2.07/2.68/0.09 and cross-domain average values of 0.84/2.01/2.72/0.09 (Diehl et al., 2024). The paper explicitly states that, in the exiD experiments, combining LoRD with multi-task fine-tuning reduces forgetting by 23.33% in ID minSFDE relative to the fine-tuning baseline (Diehl et al., 2024).
At a different level of abstraction, exiD has also been used for headway modeling. A statistical distribution paper groups exiD with highD as heterogeneous traffic under uninterrupted highway flow and fits a proposed generalized-exponential headway distribution to exiD headways after sampling at 1 Hz and filtering to $0.5$7 s. For exiD, it reports parameter estimates $0.5$8 and $0.5$9, and KL divergence/Wasserstein distance of 0.0197/0.0014 for the proposed model, outperforming six alternative distributions on that dataset (Leungbootnak et al., 5 Nov 2025).
6. Methodological caveats, interpretive tensions, and research significance
Several limitations recur in exiD-based work. Some studies use only subsets of the dataset: the merge-planning paper uses only four of seven locations and, for planning, only recording number 39, even though that recording contains many merge episodes (Mozaffari et al., 2023). Some lane-change papers exclude any sample containing multiple lane changes within the prediction horizon, retaining only trajectories with a single lane change or no lane change, which means successive lane changes, aborted maneuvers, and cut-ins are not modeled (Shi et al., 30 Dec 2025, Shi et al., 22 Sep 2025). The recurrent AoI study balances classes for training and testing but does not specify the exact balancing method, hidden-state size, or stronger split isolation in the provided description (Abouras et al., 21 Jan 2026). In adaptive decision-making, exiD contributes to the combined benchmark, but most reported metrics are pooled with highD rather than disaggregated (Huang et al., 2024).
There are also interpretive tensions that are best treated as scope differences rather than contradictions. One paper places exiD in the category of highway, uninterrupted flow for headway distribution fitting (Leungbootnak et al., 5 Nov 2025), while several others treat exiD precisely as the ramp-rich benchmark for highly interactive merging, diverging, and weaving (Mozaffari et al., 2023, Shi et al., 30 Dec 2025). This suggests that exiD can support both facility-level traffic-flow modeling and localized interaction modeling, depending on whether the unit of analysis is aggregate headway or ramp maneuver intent. A related misconception is that exiD can be substituted for highD without changing the research question; the later literature repeatedly argues against this, emphasizing that exiD’s distinctive value lies in evaluating how prediction and decision modules behave when lane semantics, merge topology, and close interactions materially affect the outcome (Mozaffari et al., 2023, Shi et al., 22 Sep 2025).
The broader significance of exiD in the recent literature is therefore methodological. It is repeatedly chosen when short-history “cold-start” prediction is operationally valuable, when left/right intent cannot be inferred from lane IDs alone, when ramp-specific topology features may improve feasibility reasoning, and when downstream planning effects must be studied on realistic merge episodes rather than on detached toy scenarios (Mozaffari et al., 2023, Shi et al., 30 Dec 2025, Mozaffari et al., 2023). A plausible implication is that exiD has become a reference benchmark for research at the interface of interaction-aware forecasting, lane-change intention recognition, prediction-for-planning evaluation, and location-shift robustness in highway entry/exit environments.