UCI-HAR

• Modality: Smartphone IMU (accelerometer + gyroscope)
• Primary Tasks: Human activity recognition from inertial sensors
• Scale: 30 subjects, 6 activity classes, 10,299 samples (2.56-second windows at 50 Hz)
• License: Public domain (UCI Machine Learning Repository)
• Access: https://archive.ics.uci.edu/ml/datasets/human+activity+recognition+using+smartphones

Summary

UCI-HAR is one of the most widely used benchmarks for smartphone-based human activity recognition. The dataset was collected from 30 volunteers (aged 19-48) wearing a Samsung Galaxy S II on the waist. Each subject performed six activities: walking, walking upstairs, walking downstairs, sitting, standing, and lying down. Tri-axial accelerometer and gyroscope signals were captured at 50 Hz, preprocessed with noise filters, and segmented into 2.56-second fixed-width sliding windows with 50% overlap. The dataset provides both raw sensor data and a 561-feature vector of time and frequency domain variables, making it accessible for both deep learning and classical ML approaches.

Reference Paper

• Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, Jorge L. Reyes-Ortiz. "A Public Domain Dataset for Human Activity Recognition Using Smartphones." ESANN, 2013. PDF

Benchmarks & Baselines

• SVM (561 features) - Accuracy: 96.0% — Anguita et al., ESANN 2013.
• DeepConvLSTM - Accuracy: ~95.8% — Ordonez & Roggen, Sensors 2016.
• 1D-CNN - Accuracy: ~96.4% — commonly reported in deep learning HAR literature.
• Standard evaluation uses the predefined 70/30 train/test split (21 train subjects, 9 test subjects).

Tooling & Ecosystem

• Available directly from the UCI ML Repository.
• Widely available in tutorial form for scikit-learn, TensorFlow, and PyTorch.
• Pre-extracted 561 features enable immediate use without signal processing.
• TensorFlow Datasets community contributions include UCI-HAR.

Known Challenges

• Only 6 activity classes; the task is considered largely solved for this label set.
• Single sensor placement (waist) limits generalizability to other body positions.
• Lab-controlled conditions do not reflect real-world deployment variability.
• Static activities (sitting vs. standing) are the primary source of confusion.
• The 561 handcrafted features may not generalize; raw signal approaches are preferred for modern work.

Cite

@inproceedings{anguita2013public,
  title     = {A Public Domain Dataset for Human Activity Recognition Using Smartphones},
  author    = {Anguita, Davide and Ghio, Alessandro and Oneto, Luca and Parra, Xavier and Reyes-Ortiz, Jorge L.},
  booktitle = {European Symposium on Artificial Neural Networks (ESANN)},
  year      = {2013}
}