# AeroIntent Dataset User Guide

AeroIntent is a simulated multimodal dataset for UAV mission intent recognition, built with Unreal Engine 5 and Cosys-AirSim. Each sequence records images, radar returns, navigation data, and annotations while a target UAV performs a particular behavior.

## 1. What the Dataset Contains

- **3 environments**: Park, Landscape (mountainous terrain), and City.
- **15 UAV configurations**: three airframe families, Drone, Inspire, and Matrice, each with five payload configurations: `no_payload`, `box`, `bomb`, `missile`, and `gun`.
- **5 mission intents** and 22 trajectory programs. The benchmark release contains **990 front-view sequences**. Additional views collected later are not included in this count.

| Intent | Directory | Behavioral Meaning |
| --- | --- | --- |
| Transit | `transit` | Passing through the target region without interacting with the target |
| Reconnaissance | `reconnaissance` | Observing the target, for example through orbiting, hovering, or figure-eight flight |
| Probing | `probe` | Approaching, probing, and withdrawing, for example through boundary probing or feint attacks |
| Strike | `strike` | Converging on the target for a terminal approach |
| Escape | `escape` | Leaving the target region, for example through high-speed withdrawal or zigzag evasion |

## 2. Files in a Sequence

A typical directory structure is:

```text
<environment>/<airframe_payload>/<intent>/<trajectory_program>/view_<view>/
```

For example: `park/Drone_no_payload/reconnaissance/orbit_reconnaissance_data_v1_6m/view_front/`.

**Each `view_*` directory is a sequence recorded from a specific viewpoint. Each frame identifier corresponds to one saved observation.**

| Data | File or Directory | Contents and Use |
| --- | --- | --- |
| RGB | `rgb/*.png` | Visible-light images showing the scene, UAV appearance, and behavior |
| Infrared (IR) | `ir/*.png` | Simulated infrared images; pixel values are not directly usable as temperatures in degrees Celsius |
| Depth | `depth/*.npz` | Raw per-pixel depth in meters; the array key is commonly `depth_m` |
| Depth preview | `depth_preview/*.png` | Depth visualizations for viewing, not a substitute for raw depth in numerical analysis |
| Echo radar | `echo/*.npy` | Per-frame radar-return point clouds, which may include both target and background returns |
| Radar point labels | `echo_groundtruth/*.txt` | The simulated object name associated with each return point; the target is usually labeled `TargetDrone` |
| IMU | `imu/imu.csv` | The target UAV's angular velocity, accelerometer output, and orientation |
| GPS | `gps/gps.csv` | The target UAV's latitude, longitude, and geographic altitude |
| 2D bounding boxes | `labels_all.txt` | The target UAV's location and size in each image, not its intent class |
| Timestamps | `timestamps.csv` | Frame identifiers and times for associating modalities and calculating time intervals |
| Metadata | `meta.json` | Viewpoint, sampling frequency, data formats, and trajectory configuration for the sequence |
| Derived videos | `rgb_video.mp4`, `ir_video.mp4`, `depth_video.mp4` | Videos for browsing and presentation; use the original per-frame files for numerical analysis |
| Radar status (optional) | `echo/echo_status.csv` | Per-frame return counts, target-hit counts, and read status; some older sequences do not contain this file |

The intent label comes from the parent directory; for example, `probe` denotes Probing. The configuration directory identifies the airframe and payload, for example `Matrice_bomb`.

## 3. Common Data Fields

### Timestamps: `timestamps.csv`

| Field | Meaning | Unit |
| --- | --- | --- |
| `frame_id` | Frame identifier within the sequence; for example, `42` corresponds to the filename stem `000042` | None |
| `sim_time_ns` | Primary timestamp for the saved frame | Nanoseconds (ns) |

The target saving rate is usually **10 Hz**, but calculate time intervals from the actual timestamps: `dt = (next_timestamp - current_timestamp) / 1e9`. Subtract the timestamps as 64-bit integers before converting the difference to seconds.

### IMU: `imu/imu.csv`

IMU data describe **the target UAV's own motion and orientation**, not the motion of the observation platform, `GroundStation`.

| Field | Meaning | Unit |
| --- | --- | --- |
| `frame_id` | Frame identifier used to associate images, timestamps, and other data | None |
| `ang_vel_x`, `ang_vel_y`, `ang_vel_z` | Angular velocity about the body X, Y, and Z axes, corresponding to roll, pitch, and yaw directions | rad/s |
| `lin_acc_x`, `lin_acc_y`, `lin_acc_z` | Accelerometer output (specific force) along the body X, Y, and Z axes | m/s^2 |
| `ori_w`, `ori_x`, `ori_y`, `ori_z` | Quaternion representing UAV orientation, stored in `[w, x, y, z]` order | None |

The body-axis convention is **X forward, Y right, and Z down**. A quaternion is neither a position nor a set of Euler angles. Accelerometer output must not be integrated directly as world-frame acceleration without accounting for orientation and gravity.

### GPS: `gps/gps.csv`

| Field | Meaning | Unit |
| --- | --- | --- |
| `frame_id` | Saved frame identifier | None |
| `lat` | Target UAV latitude | Degrees |
| `lon` | Target UAV longitude | Degrees |
| `alt` | Geographic altitude derived from the simulation's geographic origin; **not height above ground** | m |

GPS can be used to estimate the target's position and trajectory, but latitude and longitude are not metric coordinates. Convert them to a consistent local metric coordinate system before calculating distances or velocities. Adjacent saved frames may contain identical GPS readings.

### Echo Radar: `echo/*.npy`

Each frame typically contains an **`N x 6` array**, where `N` is the number of return points in that frame. The six columns are ordered as follows:

```text
x, y, z, attenuation, total_distance, reflection_count
```

| Column | Meaning | Unit |
| --- | --- | --- |
| `x`, `y`, `z` | Return-point position in the radar sensor's local coordinate frame | m |
| `attenuation` | Accumulated attenuation from simulated propagation and reflections | dB |
| `total_distance` | Accumulated ray propagation path length; with multiple reflections, this is not the straight-line distance to the target center | m |
| `reflection_count` | Number of reflections up to the return point | Count |

Labels in the matching `echo_groundtruth/*.txt` file follow the same point order and can be used to select `TargetDrone` returns. **Return points are not UAV-center positions.** The number of points can vary between frames, and a frame may contain no target returns. These six columns do not include Doppler velocity.

When `echo_status.csv` is present, its most commonly used fields are `frame_id` (frame identifier), `point_count` (total point count), `target_hit_count` (target-return point count), and `status` (read status). `status=ok` does not guarantee a target hit; check the target count or point labels as well.

### Bounding Boxes: `labels_all.txt`

The commonly used combined annotation format is:

```text
frame_id xc yc w h
```

| Field | Meaning |
| --- | --- |
| `frame_id` | Identifier of the associated frame |
| `xc`, `yc` | Normalized horizontal and vertical coordinates of the bounding-box center |
| `w`, `h` | Bounding-box width and height as fractions of the image width and height |

The image origin is at the top-left corner. For image width `W` and height `H`, the box center in pixels is `(xc * W, yc * H)`, and its size is `(w * W, h * H)`.

A line containing **only a frame identifier** indicates that the frame has no valid target bounding box; it does not necessarily mean that the target is physically absent. This format has no additional class-ID column. Obtain the five-class intent label from the sequence directory.

## 4. Key Information in `meta.json`

| Field | Meaning for Dataset Users |
| --- | --- |
| `view` | Observation viewpoint, such as `front`, `left`, `back`, or `right` |
| `frequency_hz` | Target saving rate; use timestamps for actual time intervals |
| `drone_vehicle` | Recorded target UAV, usually `TargetDrone`; IMU and GPS data come from this object |
| `camera_vehicle`, `echo_vehicle` | Observation platform carrying the cameras and radar, usually `GroundStation` |
| `modalities` | Enabled modalities; an enabled modality does not guarantee a visible target in every frame or a valid file for every observation |
| `modalities.depth` | Depth units, array key, and depth type. `DepthPlanar` measures depth along the camera's forward axis; `DepthPerspective` measures distance along each pixel ray |
| `modalities.echo_target_pose_augmentation` | Whether target-pose-based radar point augmentation was enabled; if `true`, not all points can be treated as unaugmented native returns |
| `trajectory_variation.family` | Trajectory behavior type, such as orbit reconnaissance |
| `trajectory_variation.context` | Sequence identity information, including environment, UAV configuration, and trajectory-program name |
| `trajectory_variation.sampled_params` | Configured parameters for this trajectory, such as speed, altitude, and radius; **not per-frame measured states** |

Fields such as `bbox_settings`, `tracking_settings`, and `writer_options` mainly describe how data were collected or files were generated and are generally not needed for routine data loading.

## 5. Accessing the Dataset

- [City data](https://huggingface.co/datasets/AerointentBenchmark/AeroIntent-city)
- [Landscape data](https://huggingface.co/datasets/AerointentBenchmark/AeroIntent-landscape)
- [Park data](https://huggingface.co/datasets/AerointentBenchmark/AeroIntent-park)
- [AeroIntent paper](https://arxiv.org/abs/2610.25869)

See each repository's README for download and shard-reconstruction instructions. The dataset is licensed under **CC BY-NC 4.0** for non-commercial use only. Please cite the AeroIntent paper when using the dataset in research.