Research

One published paper, one award-winning dissertation, and three directions that follow from them. The through-line is honest evaluation: build the framework that lets approaches be compared under one protocol, rather than make one approach work and call it best.

IEEE ICECET 2026, Rome, 6 to 9 July

Baseline vs Exploratory Deep Learning Video Models for Driver Drowsiness Detection: A Multi-Dataset Evaluation

Abhijith Pradeep and Wael AlSaafin, University of Warwick

Classical baselines (PERCLOS and rule-based fusion of eyelid and mouth dynamics) are compared with CNN-GRU, R(2+1)D-18 and Temporal Shift models under one protocol: subject-independent splits, 112 by 112 frames, 10 fps, 16-frame clips, across UTA-RLDD, YawDD and DROZY. Cross-domain robustness is quantified directly as the performance drop between internal and external test sets. Deep models beat the baselines, R(2+1)D-18 scored best internally, and CNN-GRU held up best on external data at a quarter of the compute, which makes it the model to deploy.

Cross-domain F1, CNN-GRU on DROZY: 0.714
Best internal F1, R(2+1)D-18 on UTA-RLDD: 0.651
Latency per clip, CNN-GRU vs R(2+1)D-18: 75 vs 305 ms

View on IEEE Xplore Read the PDF Code on GitHub

The finding worth remembering. Every model failed on YawDD, and Grad-CAM showed why: talking looks like yawning. Best-on-paper and best-in-practice came apart, and noticing that is most of the contribution.

Results

The numbers behind the finding.

UTA-RLDD, internal testDROZY, external test

F1 per model, subject-independent. The internal winner, R(2+1)D-18, is not the external winner.

Latency per clip. CNN-GRU is roughly four times faster and three times smaller, which is what makes it the one to deploy.

geometry GRU, 25,000 parameterschance or majority baseline

The deployable prototype, subject-independent across 59 subjects and 120 recordings.

Directions

Three threads, each with the cheap test that has to pass before the expensive commitment.

A drowsiness detector that ships

The training corpus was rebuilt from source at full resolution with face crops and a geometry channel: 363 recordings, 38 hours, verified at 0.9944 mean face detection. The working system is a 25,000-parameter GRU over eye and mouth aspect ratios on 12.8-second windows, subject-independent across 59 subjects, running at over 180 clips per second.

Recording-level accuracy of 70.8 percent against a 50.8 percent majority baseline, PR-AUC 0.798 against 0.508 chance. Next: an ordinal three-stage alertness head and a MobileNetV3-GRU backbone.

Phase 0.6 complete, Phase 1 next

Brain-aligned video features

Do TRIBE v2 representations, which are trained to predict brain responses, predict fatigue better than the V-JEPA 2 encoder they are built on? TRIBE's whole video pathway is V-JEPA 2, so that row is the essential control: without it a good TRIBE result could not be separated from "a better video encoder than ResNet-18".

The delta between the two is the brain-alignment effect, isolated. Phase 1 is a kill-test on one fold: 121 recordings, 20.9 hours, 150,422 forward passes, with real throughput measured on ten recordings before any of it is committed.

Phase 0 gate passed, Phase 1 running

Reservoir operation for Kerala's floods

Flood prediction is mature and well funded. Google Flood Hub gives seven-day riverine forecasts across all of India, but its own framing is "move people, not water". Control, meaning dam operation, is the gap.

A reservoir-operation model gives an asymmetric result: anticipatory release helps in prolonged-accumulation events and fails in mesoscale cloudburst events. That turns the operating question into a forecasting-classification question: can the two event types be told apart early enough to act differently? Academic contacts whose group has characterised exactly that distinction are identified, and the outreach leads with a question only they can answer.

Result in hand, outreach prepared. More under Future work

How the work gets done

Subject-independent, always

No subject appears in both train and test. The numbers come out lower than the literature's and are the ones a deployment would actually see.

Fix the confound first

Calibration was repaired before the brain-alignment work started, so any improvement can be attributed to the thing that changed.

The interesting result is the failure

Every model failing on YawDD, and Grad-CAM proving why, is worth more than another point of F1 on the easy dataset.

Cheap test before expensive commitment

A timing probe on ten recordings runs before a compute budget of one to fourteen GPU hours is committed. Hardware is bought after the audio test, not before.