Projects
Neural net drone ID system.
Reads behaviour from a live camera and scores risk in real time, entirely on the device.
Accuracy on unseen people
89%
Inference latency
<45ms
Live video
30fps
Landmarks per frame
500+
Overview
MotionTrack tracks movement, expression, gesture, posture, gaze and proximity from a live camera. A neural network turns those signals into one risk score and four clear tiers, with no cloud processing.
- Python
- PyTorch
- OpenCV
- MediaPipe
- scikit-learn
- NumPy
- Pandas
- FER
- Matplotlib
- Joblib
- yt-dlp
What it reads
Six signals, one score.
Read from every frame and fused into one risk score.
How it was built
8 steps- 1
Plan
Signals, risk tiers and speed targets.
- 2
Detect
Face, hands, pose and gaze with OpenCV and MediaPipe.
- 3
Features
Landmarks turned into 22 behavioural features.
- 4
Calibrate
A rolling baseline for each person.
- 5
Data
Labelled, balanced samples at three risk levels.
- 6
Train
A PyTorch network that scores four risk tiers.
- 7
Integrate
Async capture with inference in the background.
- 8
Test
Accuracy, latency and false positives across people and settings.
Every frame
Under 45 ms
From landmarks to a risk tier.
- 500+Landmarks
- 22Features
- 1Risk score
- 4Risk tiers
Model and demo