humanoidro/gimbaled-uav-tracking-dataset

Gimbaled UAV Tracking Dataset / 云台无人机追踪数据集

A real-world, multi-sensor dataset for active localization of a non-cooperative UAV using a two-axis gimbaled LiDAR–camera fusion system. It contains 16 flight sequences acquired with a ground vehicle platform, together with the associated tracking outputs and an RTK-based position reference.

面向非合作无人机主动定位的实测多传感数据集，采集自一套基于两轴云台 LiDAR–相机融合的地面车辆平台。包含 16 个飞行序列，并附带相应的跟踪 输出与基于 RTK 的位置参考。

The sequences span two acquisition days, sunny and overcast sky backgrounds, daylight through low-light evening imagery, and a range of target motions and scene complexity. All sequences use the same target UAV and calibrated sensing platform. The dataset therefore evaluates environmental and operational variation within this hardware configuration; a different platform requires its own sensor calibration.

这些序列跨两个采集日，覆盖晴天与多云天空背景、日间至傍晚低照度成像， 以及多种目标运动和场景复杂度。所有序列均使用同一目标无人机和同一套已标定 感知平台。因此，本数据集用于评估该硬件配置下的环境与运行条件变化；更换 平台时需要重新完成传感器标定。

---

Related Paper and Source Code / 关联论文与源代码

This dataset accompanies the paper Active Localization of Non-Cooperative UAV Based on Gimbaled LiDAR–Camera Fusion. The corresponding acquisition, processing, and deployment implementation is released separately:

本数据集对应论文《Active Localization of Non-Cooperative UAV Based on Gimbaled LiDAR–Camera Fusion》。相应的采集、处理与部署实现已单独开源：

• Repository / 仓库: https://github.com/humanoidro/gimbaled-uav-tracking
• Both the captured raw_data/ point clouds and the corresponding deskewed processed_data/ point clouds are included for inspection and evaluation.
• 数据集中同时提供采集的 raw_data/ 点云及其对应的去畸变 processed_data/ 点云，便于检查与评估。

---

Platform / 平台硬件

Component / 部件	Model / 型号	Role / 作用
Tracking LiDAR / 追踪雷达	RoboSense RS-M1 (solid-state, 78750 pts/frame)	Target ranging / 空中目标定位
Camera / 相机	Hikvision MV-CS050-10UC	UDP-triggered, YOLO detection / YOLO 检测
Gimbal IMU / 云台 IMU	WIT-motion	Gimbal attitude / 云台姿态 IMU
Vehicle IMU / 车体 IMU	WIT-motion	Vehicle attitude / 车体姿态IMU
Gimbal / 云台	2-axis (pan/tilt)	100 Hz control, 115 Hz feedback / 100 Hz 控制，115 Hz 反馈
Chassis / 底盘	CAN-bus Ackermann steering	Vehicle motion / 车辆运动
RTK	Vehicle-mounted + UAV-mounted	Reference / 参考真值

---

Sensor Extrinsics / 传感器外参

All transforms are 4×4 homogeneous matrices following the convention T_A_B maps a point from frame B into frame A (p_A = T_A_B · p_B). Machine-readable values are in script/calibration.json under sensor_extrinsics; they are identical to the source repository.

所有变换均为 4×4 齐次矩阵，遵循 T_A_B 将点从 B 坐标系变换到 A 坐标系 （p_A = T_A_B · p_B）的约定。机器可读数值见 script/calibration.json 的 sensor_extrinsics 字段。

Coordinate Frames / 坐标系定义

Symbol	Frame / 坐标系
W	World (ENU: x=East, y=North, z=Up) / 世界坐标系（ENU：x 东、y 北、z 上）
V	Vehicle (origin: ground level, centre of four wheels; x=forward, y=left, z=up) / 车辆坐标系（原点：地面四轮中心；x 前、y 左、z 上）
H	RS-Helios LiDAR (vehicle localization) / Helios 雷达（车辆定位）
G1	Gimbal yaw (pan) joint / 云台偏航关节
G2	Gimbal pitch (tilt) joint / 云台俯仰关节
L	RS-M1 LiDAR (tracking sensor) / M1 雷达（追踪传感器）
C	Camera / 相机
R	Vehicle RTK antenna / 车载 RTK 天线
U	UAV body / 无人机机体

Kinematic Chain / 运动学链

The M1 LiDAR point is transformed into the world frame by:

M1 雷达点变换到世界坐标系的完整链路为：

T_W_L = T_W_V · T_V_G1(yaw) · T_G1_G2(pitch) · T_G2_L

where T_W_V is the vehicle pose (from localization), and yaw/pitch are the real-time gimbal joint angles.

其中 T_W_V 为车辆位姿（由定位模块估计）， yaw/pitch 为实时云台关节角。

Encoder → Angle Mapping / 编码器→角度映射

θ_pan  =  0.0669 × (E_pan  − 2062.8)     [deg]
θ_tilt = −0.071  × (E_tilt − 1936.79)    [deg]

Extrinsic Values / 外参数值

The two joint frames are parameterized by rotation + fixed translation; the remaining static transforms are given as full 4×4 matrices.

两个关节坐标系以旋转 + 固定平移参数化；其余静态变换以完整 4×4 矩阵给出。

Transform / 变换	Rotation / 旋转	Translation (m) / 平移 (m)
T_V_G1(yaw)	Rz(yaw)	(0.18, 0, 0.85)
T_G1_G2(pitch)	Ry(pitch)	(0, 0, 0.062)

T_G2_L = [ 1  0  0  0.018 ]      T_V_H = [ 1  0  0  -0.17 ]      T_V_R = [ 1  0  0  0.01 ]
         [ 0  1  0  0     ]             [ 0  1  0   0    ]             [ 0  1  0 -0.22 ]
         [ 0  0  1  0.072 ]             [ 0  0  1   0.95 ]             [ 0  0  1  0.86 ]
         [ 0  0  0  1     ]             [ 0  0  0   1    ]             [ 0  0  0  1    ]

         [  0.0  -0.0192   0.9998  0.025 ]
T_G2_C = [ -1.0   0.0      0.0     0.08  ]   (rotation = ideal axis permutation
         [  0.0  -0.9998  -0.0192  0.058 ]    left-multiplied by Ry(+1.10°) fine correction)
         [  0.0   0.0      0.0     1.0   ]

Camera Intrinsics / 相机内参

fx = 2292.411   fy = 2289.128   cx = 1183.411   cy = 1038.733
k1 = -0.0827    k2 =  0.1320    k3 = -0.1129
p1 =  0.000319  p2 = -0.003229        image: 2448 × 2048

---

Dataset Structure / 数据集结构

Each flight shares the same layout:

每个 flight 目录结构一致：

flight_XX/
  ├── raw_data/<ts>/        Raw LiDAR point clouds (XYZIRT PCD), as captured
  │                          原始 LiDAR 点云（XYZIRT PCD），原始捕获
  ├── processed_data/<ts>/  Deskewed point clouds (XYZIRT PCD)
  │                          去畸变点云（XYZIRT PCD）
  ├── camera_images/        YOLO detection frames (JPEG)
  │                          YOLO 检测帧（JPEG）
  ├── imu_data/             gimbal IMU (imu_*.csv) + vehicle IMU (imu_vehicle_*.csv)
  │                          云台 IMU + 车体 IMU
  ├── servo_data/           Servo encoder positions (CSV)
  │                          舵机编码器位置（CSV）
  ├── gimbal_pose/          EKF-estimated gimbal pose (CSV)
  │                          EKF 估计的云台姿态（CSV）
  ├── tracking/             Tracking state / candidates / detection frames (CSV)
  │                          跟踪状态 / 候选 / 检测帧（CSV）
  ├── visual_tracking/      [optional] IMM-PDAF control-loop timing log (CSV)
  │                          [可选] IMM-PDAF 控制环时序日志（CSV）
  └── drone_rtk.csv         UAV RTK reference (GGA sentences)
                             无人机 RTK 参考（GGA 语句）

The tracking/ CSV files and processed_data/ point clouds are released system outputs accompanying the sensor recordings.

tracking/ 下的 CSV 与 processed_data/ 点云是随传感器记录一并发布的 系统输出。

visual_tracking/ is present only in some flights for latency analysis.

visual_tracking/ 仅部分 flight 存在用于耗时分析。

---

File Formats and Time Conventions / 文件格式与时间约定

• LiDAR files use binary PCD with fields x y z intensity ring timestamp. Coordinates are in metres. The absolute point time in seconds is TIMESTAMP_OFFSET + timestamp, where TIMESTAMP_OFFSET is stored in each PCD header.
• 点云采用二进制 PCD，字段为 x y z intensity ring timestamp，坐标单位为米。点的绝对秒级时间为 PCD 头部的 TIMESTAMP_OFFSET + timestamp。
• CSV columns ending in _us are timestamps or durations in microseconds. In drone_rtk.csv, the second field is a Unix timestamp in milliseconds; the file contains NMEA-derived GGA/GST records.
• CSV 中以 _us 结尾的字段表示微秒时间戳或时长。drone_rtk.csv 的第二列为 Unix 毫秒时间戳，文件保存由 NMEA 解析得到的 GGA/GST 记录。
• Coordinate-frame definitions and all evaluation transforms are recorded in script/calibration.json. CSV headers are self-describing, and script/evaluate_accuracy.py is the reference parser for the RTK-based accuracy evaluation.
• 坐标系定义与评估变换均记录在 script/calibration.json 中。CSV 首行为字段 名称；RTK 精度评估的参考解析方式见 script/evaluate_accuracy.py。

---

Flights / 飞行序列

Acquisition days: 2026-07-16 (flights 01–05, 08, 10, 12, 13) and 2026-07-23 (flights 06, 07, 09, 11, 14, 15, 16).

采集日期：2026-07-16（flights 01–05、08、10、12、13）与 2026-07-23（flights 06、07、09、11、14、15、16）。

The collection includes sunny and overcast sky backgrounds and extends from normal daylight to lower-light evening images.

采集条件包含晴天和多云天空背景，并覆盖正常日间光照至傍晚低照度图像。

Flight	Date	Trajectory / 轨迹	Difficulty	LiDAR frames	Camera frames
flight_01	2026-07-16	Stable hover / 稳定悬停	easy	744	937
flight_02	2026-07-16	Low-dynamic hover w/ vertical drift / 低动态悬停含垂直漂移	easy	828	1002
flight_03	2026-07-16	Vertical ascent/descent + hover / 垂直升降及两端悬停	medium	707	868
flight_04	2026-07-16	Lateral left-to-right / 左向右横移	medium	494	662
flight_05	2026-07-16	Lateral right-to-left / 右向左横移	medium	537	777
flight_06	2026-07-23	Radial approach then retreat / 径向接近后远离	medium	998	1171
flight_07	2026-07-23	Longitudinal/radial round-trip / 纵向径向往返	medium	668	818
flight_08	2026-07-16	Wide-arc continuous turn / 宽弧线连续转向	hard	731	912
flight_09	2026-07-23	Figure-8 / continuous turns / 8 字连续左右转向	hard	934	1233
flight_10	2026-07-16	Close-range lateral fly-by / 近距横向掠过	hard	535	676
flight_11	2026-07-23	Lateral crossing / 横向飞越	hard	462	913
flight_12	2026-07-16	Natural track-loss & re-acquisition / 自然失跟与重捕获	hard	778	922
flight_13	2026-07-16	Close-range complex background tracking / 近距复杂背景连续跟踪	hard	681	999
flight_14	2026-07-23	Complex background lateral/slow-arc / 严格复杂背景横移缓弧	hard	1161	1353
flight_15	2026-07-23	Complex background repeat / 复杂背景重复	hard	521	681
flight_16	2026-07-23	Long-range hover & slow lateral / 远距悬停及缓慢横移	hard	1483	1666

Totals: 12,262 LiDAR frames / 15,590 camera frames, ≈ 42 GB. 合计：12,262 帧点云 / 15,590 帧图像，约 42 GB。

---

Reproducing the Accuracy Evaluation / 复现精度评估

Python 3.10 or newer is recommended. Install the two runtime dependencies from the supplied requirements file:

建议使用 Python 3.10 或更高版本，并通过随附文件安装两个运行依赖：

python3 -m pip install -r requirements.txt

Then run the supplied evaluation and visualization scripts:

随后运行评估与可视化脚本：

# Accuracy metrics (per-flight RMSE, sensitivity by range/azimuth/elevation,
#                   bootstrap CI, day-to-day comparison)
# 精度指标（逐 flight RMSE、距离/方位/俯仰敏感性、自助法置信区间、跨日对比）
python3 script/evaluate_accuracy.py .

# Interactive 3D trajectory visualization (HTML)
# 交互式 3D 轨迹可视化（HTML）
python3 script/generate_trajectory_html.py .

Results are written to script/output/. A pre-generated copy is already included for reference.

结果写入 script/output/，目录内已附一份预生成结果供参考。

Output	Description / 说明
accuracy_by_flight.csv	Per-flight accuracy / 各航次精度
accuracy_summary.json	Aggregate metrics and evaluation rules / 汇总指标与评估规则
paired_errors.csv	Paired tracker–RTK samples / 跟踪器与 RTK 配对样本
range_azimuth_elevation_sensitivity.csv	Range/azimuth/elevation analysis / 距离、方位与俯仰分解
uav_lever_arm_sensitivity.csv	UAV RTK lever-arm sensitivity / 无人机 RTK 杆臂敏感性
trajectories.html	Offline interactive trajectories / 离线交互轨迹

The included result evaluates all 16 flights and 9,442 paired tracker–RTK samples. The unweighted mean of the per-flight 3D RMSE values is 0.528 m, the corresponding mean per-flight P95 is 0.849 m, and the all-sample (micro) 3D RMSE is 0.566 m. Selection, interpolation, and aggregation rules are stored with the result in accuracy_summary.json.

随附结果覆盖全部 16 个航次和 9,442 个跟踪器–RTK 配对样本。逐航次等权 汇总的三维 RMSE 为 0.528 m，对应的逐航次 P95 均值为 0.849 m， 全部样本汇总（micro）的三维 RMSE 为 0.566 m。样本选择、插值及汇总 规则均随结果记录在 accuracy_summary.json 中。

---

License and Citation / 许可与引用

The dataset is released under CC0 1.0 Universal; see LICENSE. No attribution is legally required.

本数据集采用 CC0 1.0 Universal 发布，完整文本见 LICENSE。法律上不强制署名。