Driver Monitoring
Real-time four-class driver-behaviour recognition from monocular RGB video.
My role
MSc Artificial Intelligence Researcher
Overview
What I built.
MSc Artificial Intelligence dissertation exploring lightweight spatiotemporal deep-learning pipelines for detecting normal, aggressive, distracted and drowsy driving behaviours in real time.
Challenge
The problem.
Recognise behaviour from short video sequences rather than isolated frames while keeping inference efficient enough for a real-time prototype.
Contribution
What I did.
Prepared balanced 8-second video clips and built preprocessing for sampled RGB frame sequences.
Used a MobileNetV2 visual backbone with temporal alternatives including BiLSTM, Transformer and temporal-convolution approaches.
Evaluated models using accuracy, precision, recall, F1, macro ROC-AUC, mAP, calibration/ECE and confusion-style error analysis.
Investigated class imbalance, dominant-class behaviour, prediction instability and misclassification patterns.
Built rolling-buffer live inference with temporal smoothing and latency/FPS testing.
Architecture
System flow.
Highlights
Outcome
A research pipeline connecting comparative deep-learning experimentation to a working real-time inference prototype.