OPPORTUNITY¶

Modality: Wearable IMUs, body-worn accelerometers, ambient sensors, location beacons
Primary Tasks: Context-aware activity recognition, sensor fusion, domain adaptation
Scale: 4 participants, ~6 hours per subject, 72 sensor channels at 30-100 Hz
License: Creative Commons Attribution-NonCommercial 4.0
Access: https://archive.ics.uci.edu/ml/datasets/OPPORTUNITY+Activity+Recognition

Summary¶

The OPPORTUNITY dataset captures rich multi-modal sensor streams during scripted and unscripted daily-living scenarios. Its dense annotation (gestures, locomotion, high-level activities) and multi-device fusion make it ideal for research on context recognition and missing-sensor robustness.

Reference Paper¶

Daniel Roggen et al. "Collecting Complex Activity Datasets in Highly Rich Networked Sensor Environments." INSS, 2010. PDF

Benchmarks & Baselines¶

Hierarchical Conditional Random Fields - F1: 78.7 for locomotion; Roggen et al., 2010.
DeepConvLSTM - F1: 86.0 (locomotion), 70.1 (gestures); Ordonez & Roggen, IJCNN 2016.
Common protocol: preprocess into sliding windows (1s, 50% overlap) and report macro F1 for locomotion and gestures separately.

Tooling & Ecosystem¶

Opportunity Processing Toolkit includes normalization and LOSO splits.
TensorFlow HAR examples.
Feature extraction via tsfresh and Pandas for classical baselines.

Known Challenges¶

Sensors operate at different frequencies; alignment and interpolation are mandatory.
High class imbalance (gestures vs. null activity) necessitates rebalancing strategies.
Dataset includes missing segments; handle NaNs carefully or impute.

Cite¶

@inproceedings{roggen2010opportunity,
  title     = {Collecting Complex Activity Datasets in Highly Rich Networked Sensor Environments},
  author    = {Roggen, Daniel and Calatroni, Alberto and Rossi, Mirco and others},
  booktitle = {International Conference on Networked Sensing Systems},
  year      = {2010}
}