Initializing Core Systems
AI & Machine Learning

Stress & Fatigue Detection via Facial Expression Recognition

Deep learning facial analysis using ViT and BEiT transformers to identify stress metrics. Official implementation of the ICCSA 2025 paper.

Model Architecture

ViT & BEiT

Benchmark Dataset

FERPlus Dataset

Core Focus

Expression Recognition

Dataset Scale

35k Images

The Bottleneck

Problem Statement

Fatigue and stress detection in drivers or high-concentration personnel is critical to prevent accidents, yet traditional CNN models suffer from low accuracy under variable lighting or subtle facial expressions.

The Architecture

Implemented Solution

Developed an end-to-end deep learning system leveraging transformer-based architectures (ViT, BEiT) trained on the FERPlus dataset, improving model accuracy and establishing robust facial fatigue diagnostics.

System Processing Pipeline

1

Preprocessing

Face alignment, scaling, and contrast equalization using OpenCV.

2

Feature Extraction

Fine-tuning Vision Transformer (ViT) and BEiT backbones.

3

Emotion Classification

Multi-class classification head mapping fatigue metrics.

4

Diagnostics output

Real-time telemetry exposing stress parameters.

Telemetry & Execution Telemetry

LIVE NODE MONITOR
ACTIVE: 8 NODESGPU: 92% AVG
INFERENCE QUEUE

11,204 QPS

VRAM CONSUMPTION

62.4 GB / 80GB

GRADIENT BUBBLES

0.02ms sync

Pipeline Technology Stack

PyTorchVision TransformersBEiTOpenCVPythonscikit-learnStreamlit

Technical Challenges & Mitigations

Challenge

Vision Transformers require large datasets and suffer from overfitting on small facial expression sets.

Mitigation

Employed advanced data augmentation techniques, transfer learning from pre-trained ImageNet weights, and regularized dropout schedules.

Lessons Learned

  • Self-attention layers in ViT capture holistic expression features better than localized CNN filters on emotion-centric tasks.
  • Dynamic data cleaning is crucial to resolve labeling discrepancies in crowd-sourced datasets like FERPlus.

Future Optimizations

  • Incorporate temporal sequence models (LSTMs/Transformers) to analyze video streams rather than single frames.
  • Deploy lightweight MobileViT variants for deployment on mobile and automotive embedded units.