Mining smartphone sensor data with python
summary
Collects smartphone accelerometer CSVs with an open-sourced Android app, extracts windows (5–20s), computes statistical/time‑series/signal features including a magnitude vector, and trains scikit‑learn models. A dummy baseline sits near 20% accuracy while a random forest reaches 98% accuracy on the hacked single‑user dataset. Notes real issues: phone orientation, sampling rate, battery drain, noisy public data, device variability and poor cross‑user generalisation; code and iPython notebook live on the GitHub repo.
AI-written summary · about
chapters (21)
- 0:00Smartphone sensor research intro
- 3:21Android data collection app
- 4:26Accelerometer data structure
- 6:53Phone orientation challenges
- 7:40Axes to magnitude conversion
- 10:30Activity classification overview
- 12:24Sliding window for time series
- 13:48Statistical features extraction
- 16:42Training classifiers with scikit-learn
- 18:49Model performance evaluation
- 20:33Real-world sensor complications
- 23:04Building activity classifier
- 23:46Sensor data for smoking cessation
- 24:52Q&A window sizing clustering
- 25:40Q&A data variability recognition
- 27:34Q&A binary classification edge cases
- 28:24Q&A time series modeling
- 29:34Q&A data balancing training
- 30:44Q&A on-device ML
- 31:45Q&A multi-sensor efficiency
- 33:10Q&A feature engineering
related talks
22mRunning, walking, sitting or biking? - Motion prediction with acceleration and rotation
27mDissecting ActivityRecognition with Multi-Sensor Simulation and Its Practical Deployment
29mIOT, timeseries and prediction with Android, Cassandra and Spark
33mAnalytics in the age of the Internet of Things #bbuzz
44mQuantified Self: Analyzing the Big Data of our Daily Life
08mSenseCollect: We Need Efficient Ways to Collect On-body Sensor-based Human Activity Data!
26mMaking Sense The Road to mobile Awareness
52mMuvr