KARD - Kinect Activity Recognition Dataset

KARD contains 18 Activities. Each activity is performed 3 times by 10 different subjects.

1 Horizontal arm wave
2 High arm wave
3 Two hand wave
4 Catch Cap
5 High throw
6 Draw X
7 Draw Tick
8 Toss Paper
9 Forward Kick
10 Side Kick
11 Take Umbrella
12 Bend
13 Hand Clap
14 Walk
15 Phone Call
16 Drink
17 Sit down
18 Stand up

In total, you have 4 (files) x 18 (activities) x 3 (repetitions) x 10 (subjects), that is 2160 files.
Each filename is in the form aA_sS_eN_string where A is a two-digit actionID and S is a two-digit subjectID for the N-th repetition.

The string parameter depends on the the type of provided information:

- depthmaps.txt: depth map,
- .mp4: 640x480 RGB video,
- realworld.txt: joints position in real world coordinates,
- screen.txt: joints position in screen coordinates and depth value.

For example, the file a04_s03_e02_realworld.txt contains the skeleton joints position in real world coordinates for the second repetition of the action #4 performed by the subject #3.
The files containing the skeleton coordinates (realworld.txt and screen.txt) list the 15 joints in consecutive blocks, one for each frame.

line 1 Head
line 2 Neck
line 3 Right Shoulder
line 4 Right Elbow
line 5 Right Hand
line 6 Left Shoulder
line 7 Left Elbow
line 8 Left Hand
line 9 Torso
line 10 Right Hip
line 11 Right Knee
line 12 Right Foot
line 13 Left Hip
line 14 Left Knee
line 15 Left Foot

Each file contains 15xF lines, where F is the number of frames for that sequence, and each line reports three numbers: real world coordinates (x, y, z) for realworld.txt, or screen coordinates and depth value (u, v, depth) for screen.txt.

The dataset is made of 540 sequences for about a total of 1 hour of videos captured at a resolution of 640x480 pixels at 30fps. Uncompressed frame images are also available on request.

If you use this dataset, please cite the following paper:

Human Activity Recognition Process Using 3-D Posture Data. S. Gaglio, G. Lo Re, M. Morana. In IEEE Transactions on Human-Machine Systems. 2014 doi: 10.1109/THMS.2014.2377111

KARD数据集包含18项活动。每一项活动由10位不同的受试者分别重复3次进行。

活动列表如下：
1. 水平手臂波浪
2. 高手臂波浪
3. 双手波浪
4. 捕捉帽子
5. 高抛
6. 画X
7. 画勾
8. 扔纸
9. 正面踢
10. 侧面踢
11. 拿伞
12. 弯腰
13. 手掌拍击
14. 步行
15. 电话通话
16. 喝饮料
17. 坐下
18. 站立

总计，您将拥有4（文件）x 18（活动）x 3（重复次数）x 10（受试者），即2160个文件。

每个文件名遵循aA_sS_eN_string的格式，其中A为两位数的动作ID，S为两位数的受试者ID，N为第N次重复。

字符串参数取决于提供信息的类型：
- depthmaps.txt：深度图
- .mp4：640x480 RGB视频
- realworld.txt：关节在真实世界坐标系中的位置
- screen.txt：关节在屏幕坐标系中的位置和深度值

例如，文件a04_s03_e02_realworld.txt包含了动作#4由受试者#3进行的第二次重复的真实世界坐标系中的骨骼关节位置。

包含骨骼坐标的文件（realworld.txt和screen.txt）按连续块列出15个关节，每个块对应一帧。

行 1：头部
行 2：颈部
行 3：右肩
行 4：右肘
行 5：右手
行 6：左肩
行 7：左肘
行 8：左手
行 9：躯干
行 10：右髋
行 11：右膝
行 12：右脚
行 13：左髋
行 14：左膝
行 15：左脚

每个文件包含15xF行，其中F是该序列的帧数，每行报告三个数字：realworld.txt中的真实世界坐标（x, y, z），或screen.txt中的屏幕坐标和深度值（u, v, depth）。

该数据集由540个序列组成，大约总时长为1小时的视频，以640x480像素的分辨率在30fps下捕获。如需请求，可提供未压缩的帧图像。

若使用本数据集，请引用以下论文：
使用3-D姿态数据的人体活动识别过程。S. Gaglio，G. Lo Re，M. Morana。发表于IEEE Transactions on Human-Machine Systems，2014年，doi: 10.1109/THMS.2014.2377111