Classroom Behavior Detection System — YOLOv8x · 74.85% mAP@0.5

type

Computer Vision

status

Active Research

year

2025

role

Research Assistant

01 —

System Architecture · 3D View

02 —

Architecture Diagram

SCB-05 Dataset

Roboflow Annotated

↓

Augmentation

Resize · Flip · Mosaic

↓

YOLOv8x

A100 · Colab Pro

↓

Training Loop

YAML Config · GPU

↓

Evaluation

mAP@0.5 = 74.85%

Grad-CAM

Explainability

↓

Tiled Inference

CCTV · RTSP Feed

Detection Output

Bounding Boxes · Live

03 —

Training Results & Proof

74.85%

mAP@0.5

Best epoch 24

94.8%

Sleeping AP

Highest class

88.8%

Phone AP

Using phone class

150

Epochs

A100 · Colab Pro

Training & Validation Curves · 150 epochs

Click to expand

Per-Class AP

Sleeping94.8%

Using Phone88.8%

Reading68.8%

Hand Raising66.3%

Writing49.5%

Overall mAP@0.5

74.85%

Best at epoch 24 / 150

Precision-Recall Curve

04 —

Screenshots & Output

terminal

$ yolo detect train data=scb05.yaml model=yolov8x.pt epochs=50

Device: A100 SXM4 80GB · Google Colab Pro

Epoch 1/50: box_loss=4.12 cls_loss=3.87

Epoch 25/50: mAP50=0.61 Precision=0.68

Epoch 40/50: mAP50=0.72 Precision=0.78

Epoch 50/50: mAP50=0.7485 Precision=0.812

✓ Best model saved → runs/detect/train/weights/best.pt

✓ Sleeping class AP: 0.935 — highest per-class score

$ python tiled_inference.py --source classroom.mp4

✓ Tiled inference complete → output_tiled.mp4

Training Output

YOLOv8x Colab A100 training logs

Class mAP Scores

Overall mAP@0.575%

Sleeping94%

Precision81%

Recall74%

Using Phone68%

Class mAP Scores

Per-class detection performance

YAML Config

{
# scb05.yaml
path: ./SCB-05
nc: 11
names: [sleeping, writing, phone, reading...]
# model=yolov8x.pt epochs=50 best_mAP50=0.7485
}

Training Config

YAML training configuration

Project Structure

📁 classroom-behavior-detection/

├─ YOLOv8x_SCB05.ipynb Colab notebook

├─ tiled_inference.py CCTV pipeline

├─ gradcam_viz.py Explainability

├─ SCB-05/ Roboflow export

├─ scb05.yaml Dataset config

└─ runs/detect/weights/ best.pt

Dataset Structure

SCB-05 Roboflow directory layout

05 —

What I Built

›

Trained YOLOv8x on the SCB-05 dataset using Google Colab Pro with A100 GPU, achieving 74.85% mAP@0.5 across 11+ behavioral classes.

›

Achieved 93.5% detection accuracy on sleeping behavior — the highest per-class precision in the model.

›

Built tiled inference pipeline for live CCTV/RTSP feed processing, enabling real-time classroom monitoring.

›

Curated and annotated multi-class dataset via Roboflow with rigorous quality control across hundreds of images.

›

Configured reproducible YAML-based training pipelines with systematic hyperparameter optimization (batch size, image size, epochs).

›

Integrated Grad-CAM explainability to validate that the model focuses on posture and body position rather than facial features.

›

Designed architecture for deployment on resource-constrained campus hardware with streaming video input.

06 —

Project Insights

✎Personal Notes & Learnings

Markdown Editor

Live Preview

Research Context

This is my primary research assistantship project at Lawrence Technological University — a production-grade classroom behavior detection system trained on the SCB-05 dataset.

Key Results

74.85% mAP@0.5 overall across 11+ behavioral classes
93.5% accuracy on sleeping detection — highest per-class precision
Trained on Google Colab Pro (A100 GPU) for fast iteration

Technical Highlights

YOLOv8x (extra-large model) chosen for maximum accuracy over speed
Roboflow used for dataset curation, annotation, and augmentation pipeline
Tiled inference enables deployment on wide-angle CCTV footage without losing small-object detection
Grad-CAM confirmed model attends to posture/body position, not faces — responsible AI validated

Next Steps

Streamlit analytics dashboard for live monitoring
Real-time RTSP stream integration for campus deployment
Paper submission to IJSRST (Under Review)

✓ Insights saved locally

← PreviousResume Intelligence Engine

All Projects

Next →Buffer Overflow Attack Lab