Human Activity Recognition (HAR) Datasets

We compiled multiple human activity recognition (HAR) datasets, each collected using IMU sensors. These datasets cover four activity states: walking, ascending stairs, descending stairs, and sitting. Further details can be found in the paper Time Series Adaptation Network for Sensor-based Human Activity Recognition.

📂 Repository: https://github.com/jiehu01/Human-Action-Recognition-Sensor-Dataset

1. PAMAP2 Dataset [1]

Contains 18 types of physical activities from 9 subjects. Includes accelerometers on arm, chest, and ankle, plus heart rate monitor. Our usage: accelerometer data from the arm sensor. Sampling frequency: 100 Hz.

2. WISDM Dataset [2]

Collected from 51 users with a smartphone in the pants pocket. Activities: walking, jogging, stairs, standing, sitting. Our usage: accelerometer data only. Sampling frequency: 20 Hz.

3. UCI HAR Dataset [3]

Includes 30 subjects (ages 19–48) performing six activities. Data recorded at waist or free position. In addition to IMU, video recordings were collected for manual annotation. Our usage: accelerometer data only. Sampling frequency: 50 Hz.

Data Organization

Only accelerometer data are used. Data are organized as <X.npy, Y.npy> pairs:

X.npy: shape [num_samples, 128, 3], stores sensor data
Y.npy: shape [num_samples], stores labels

Example: X.npy = [5,128,3], Y.npy = [5] → 5 samples per Raw Segment.

Data Preprocessing

- Selected activities: walking, upstairs, downstairs, sitting.
- Unified sampling frequency to 50 Hz (linear interpolation).
- Applied sliding window (length 128, 2.56s), non-overlapping.
- Raw segments divided into fixed-size samples.

Processed Data Statistics

Dataset	Users	Raw Segments	Samples	Samples per Activity (Walking, Upstairs, Downstairs, Sitting)
UCIHAR	30	616	3374	893, 809, 746, 926
WISDM	51	323	13658	8258, 2341, 1901, 1158
PAMAP2	9	361	4727	1584, 905, 808, 1430

Labels are represented by numbers in the same order as the table above.

Recommended Citation

Wen S, Chen Y, Ma Y, et al. Time series adaptation network for sensor-based cross domain human activity recognition. IJCNN 2023.

References

Reiss A, Stricker D. Introducing a new benchmarked dataset for activity monitoring. ISWC 2012.
Kwapisz J R, Weiss G M, Moore S A. Activity recognition using cell phone accelerometers. ACM SigKDD Explorations 2011.
Anguita D, Ghio A, Oneto L, et al. A public domain dataset for human activity recognition using smartphones. ESANN 2013.
Wen S, Chen Y, Ma Y, et al. Time series adaptation network for sensor-based cross domain human activity recognition. IJCNN 2023.