GitHub项目推荐 | Comma.ai发布无人驾驶数据集comma2k19
GitHub项目推荐 | Comma.ai发布无人驾驶数据集comma2k19
项目地址：https://github.com/commaai/comma2k19
comma.ai 发布了 comma2k19, 这是加利福尼亚280高速公路上超过33小时通勤的数据集。 这意味着在加利福尼亚州圣何塞和旧金山之间20公里的高速公路上行驶了2019段，每段1分钟。 comma2k19是一个完全可重现且可扩展的数据集。 数据采用comma EONs收集，其传感器类似于任何现代智能手机，包括道路相机，手机GPS，温度计和9轴IMU。 此外，EON还使用comma grey panda捕获原始GNSS测量值和汽车发送的所有CAN数据。
出版刊物
有关此数据集的详细说明，请参阅我们的 论文。 如果您在研究中使用comma2k19或Laika，请考虑引用。
@misc{1812.05752,
Author = {Harald Schafer and Eder Santana and Andrew Haden and Riccardo Biasini},
Title = {A Commute in Data: The comma2k19 Dataset},
Year = {2018},
Eprint = {arXiv:1812.05752},
}下载
数据集有100GB这么大，所以将以10GB为一个数据包进行下载，链接 戳此
示例代码
在这份实验报告中有一个用于实验的示例数据段。还有一些笔记里带有一些示例代码，包括位置基准。这些代码只在python 2.x和ubuntu 16.04上进行了测试。 如果您尚未安装相关软件包，请运行命令 pip install -r requirements_examples.txt。 这些示例包含1分钟的示例和一些示例笔记。
- processed_readers：数据读取和绘图的一些示例
- position_benchmarks：运行用于评估修复质量的位置基准的示例
- raw_readers：使用openpilot_tools的示例
有关原始GNSS的示例，请查看 Laika
数据集结构
目录结构
数据被分成10块，每一块大约200分钟的车程。数据集的1-2块是RAV4，其余的是civic。RAV4的dongle_id是b0c9d2329ad1606b, civic的dongle_id是99c94dc769b5d96e。
Dataset_chunk_n
|
+-- route_id (dongle_id|start_time)
|
+-- segment_number
|
+-- preview.png (first frame video)
+-- raw_log.bz2 (raw capnp log, can be read with openpilot-tools: logreader)
+-- video.hevc (video file, can be read with openpilot-tools: framereader)
+-- processed_log/ (processed logs as numpy arrays, see format for details)
+-- global_pos/ (global poses of camera as numpy arrays, see format for details)日志格式
processed_log控制器中的每个日志类型都包含2个numpy数组。 使用系统设备的引导时间和值数组的时间戳数组（以秒为单位）。
processed_log
|
+--IMU ([forward, right, down])
| |
| +--acceleration: (m^2/s)
| +--gyro_uncalibrated (rad/s)
| +--gyro_bias: android gyro bias estimate (rad/s)
| +--gyro: with android bias correction (rad/s)
| +--magnetic_uncalibrated: (T)
| +--magnetic: with android calibration(T)
|
+--CAN data:
| |
| +--car_speed (m/s)
| +--steering_angle (deg)
| +--wheel_speeds: [front_left, front_right, rear_left, rear_right] (m/s)
| +--radar: [forward distance (m),
| | left distance (m),
| | nan,
| | nan,
| | address,
| | new_track (bool)]
| +--raw CAN: This not stored as a value array but as three seperate arrays [src, address, data]
|
+--GNSS
|
+--live_gnss_qcom: [latitude (deg),
| longitude (deg),
| speed (m/s),
| utc_timestamp (s),
| altitude (m),
| bearing (deg)]
+--live_gnss_ublox: [latitude (deg),
| longitude (deg),
| speed (m/s),
| utc_timestamp (s),
| altitude (m),
| bearing (deg)]
|
+--raw_gnss_qcom: every row represents a measurement
| of 1 sattelite at 1 epoch can easily
| be manipulated with laika.
| [prn (nmea_id, see laika),
| week of gps_time of reception (gps_week),
| time pf week of gps_time of reception (s),
| nan,
| pseudorange (m),
| pseudorange_std (m),
| pseudorange_rate (m/s),
| pseudorange_rate_std (m/s)]
+--raw_gnss_ublox: every row represents a measurement
of 1 sattelite at 1 epoch can easily
be manipulated with laika.
[prn (nmea_id, see laika),
week of gps_time of reception (gps_week),
time pf week of gps_time of reception (s),
GLONASS channel number (-7..6) nan if not GLONASS,
pseudorange (m),
pseudorange_std (m),
pseudorange_rate (m/s),
pseudorange_rate_std (m/s)]姿势格式
存储摄像机的姿势和视频的每帧的时间戳
如下：
frame_times: timestamps of video frames in boot time (s)
frame_gps_times: timestamps of video frames in gps_time: ([gps week (weeks), time-of-week (s)])
frame_positions: global positions in ECEF of camera(m)
frame_velocities: global velocity in ECEF of camera (m/s)
frame_orientations: global orientations as quaternion needed to
rotate from ECEF frame to local camera frame
defined as [forward, right, down] (hamilton quaternion!!!!)联系
有任何问题、疑虑或者建议，请联系harald@comma.ai
项目地址：https://github.com/commaai/comma2k19