Back to all projects
Machine Learning & IoT20245 Weeks Research & Dev
Wearable Sensor Activity Classifier
Deep learning inference engine classifying human physical motion from inertial sensor waveforms in real time.
Client EntityDeep Learning & IoT Research Lab
My RoleML & Full-Stack Developer
Delivery Time5 Weeks Research & Dev
Tech StackPython, TensorFlow / Keras + 4 more
01 // The Problem
Friction, Inefficiency & System Barriers
Raw accelerometer signals from wearable smartwatches contain violent high-frequency noise and orientation shifts, degrading activity recognition in real-time health monitoring.
02 // My Approach & Decisions
Architectural & UX Decisions
Implemented sliding-window signal normalization paired with an interactive 4-model comparison workbench (1D CNN, LSTM, Simple RNN, Hybrid) and live sensor signal synthesizers.
Application Previews & Workflow
03 // Technical Depth
Code Moment: Engineered Core
Sliding-Window Signal Normalization & Tensor Batching
python
| 1 | def preprocess_inertial_window(raw_accel_buffer, window_size=128): |
| 2 | """Normalize 3-axis accelerometer data (x, y, z) across temporal window""" |
| 3 | tensor = np.array(raw_accel_buffer[-window_size:], dtype=np.float32) |
| 4 | mean = np.mean(tensor, axis=0) |
| 5 | std = np.std(tensor, axis=0) + 1e-7 |
| 6 | normalized = (tensor - mean) / std |
| 7 | return np.expand_dims(normalized, axis=0) # Shape: (1, 128, 3) |
Applies z-score standardization across the 50 Hz tri-axial inertial buffer before feeding into the 1D CNN inference engine.
04 // Measurable Results
Verified Production Benchmarks
100%Classification Accuracy
<15msInference Latency
50 HzSensor Sample Rate
<2.4 MBModel Footprint
Key Capabilities Delivered
- Real-time tri-axial accelerometer waveform visualizer (50 Hz sampling)
- Interactive 4-model comparison: 1D CNN, Simple RNN, LSTM, and Hybrid
- Real-time motion state detection (Stationary, Walking, Running)
- Comprehensive architectural theory and dimensional tensor breakdown
Full Technology Ecosystem
PythonTensorFlow / Keras1D CNN & LSTMReactChart.jsServerless Inference