UNDERSTAND HUMAN MOVEMENT USING SENSOR DATA
An advanced machine learning system trained on the official UCI Human Activity Recognition dataset, analyzing 561-dimensional smartphone sensor readings to classify activities with high accuracy.
ACTIVITY PREDICTOR
Describe the movement using simple observable characteristics to find and analyze real dataset observations.
Questionnaire
Adjust the parameters to select matching smartphone recordings from the real UCI test set.
Analysis Result
Configure the physical descriptors and click Analyze Activity to run machine learning predictions on matching records.
ANALYZE REAL UCI RECORDING
Directly load individual 561-dimensional rows from the official test dataset (`X_test`) to evaluate model predictions against ground truth (`y_test`).
Sample Selection
Choose a test sample between index 1 and 2,947 to inspect. The backend loads the exact feature vector from `X_test.txt` and queries the ML model.
Verification Results
Select a sample index and click Analyze Sensor Recording to inspect the row data and model decision.
Top Sample Probabilities
LIVE SENSOR PREDICTOR
Real-time smartphone activity classification using a 128-sample sliding window, 50 Hz resampling, 3D gravity alignment, digital signal processing, and a trained ML model.
Sensor Access
This mode uses your device's motion sensors via the browser DeviceMotion API. Open this page on your smartphone and grant sensor permission to begin.
Sensor Status & Controls
Current Activity
Waiting for sensor access. Grant permission and hold your phone to begin classifying activity.
š¶ WALKING
Probability Distribution
Activity History
- No activities logged yet.
MODEL DASHBOARD
Real-time statistics and training evaluation metrics saved in the system metadata.
Model Visualizations
Confusion matrix for the selected model. Displays counts of correct classifications vs misclassifications per class.
Accuracy comparison across all tested machine learning algorithms on the official test set.
Counts of observation samples representing each activity in the training dataset.
HOW IT WORKS
Understanding the signal processing and machine learning pipeline.
Smartphone Sensors
Triaxial raw accelerometer (linear acceleration) and gyroscope (angular velocity) signals captured at 50 Hz.
Signal Processing
Noise filtering and gravity/body acceleration separation using low-pass Butterworth filters (0.3 Hz corner).
Feature Engineering
Extraction of 561 time and frequency domain variables (mean, standard deviation, entropy, autoregression, etc.).
Trained ML Model
Multi-class classification using an optimized Support Vector Machine (SVM) running on the 561 features.
Activity Prediction
Output of final predicted activity name along with exact class probability distribution.
š¬ Scientific Note on Sensor Compatibility
Important: This web application uses real recorded smartphone sensor data from the official UCI dataset. It does not pull live readings from your phone's browser.
To support live phone integration, raw values (Acc X/Y/Z, Gyro X/Y/Z) cannot be directly fed into the model. They must first undergo windowing and identical digital signal processing (DSP) to extract the exact 561 features on which the SVM was trained.
THE SIX CLASSIFIED ACTIVITIES
Review characteristics of each classified physical movement.
Walking
Dynamic cyclic gait on flat ground. Marked by regular body acceleration peaks and moderate rotation velocity.
Walking Upstairs
Dynamic traversal upward. Shows higher body acceleration intensity and moderate orientation deviation.
Walking Downstairs
Downward descent with high impact peaks. Typically exhibits maximum acceleration variation (low stability).
Sitting
Static seated position. Minimal body acceleration, highly stable gravity alignment along vertical axes.
Standing
Static standing position. Very low motion amplitude, stable vertical alignment of the phone.
Lying
Static horizontal orientation. Gravity vector shifts completely along horizontal axes of the phone.