BITS Pilani  ·  M.Tech Software Engineering  ·  SEZG628T

ROAD
QUALITY
MAP

Every road segment, scored by AI. Dashcam footage transformed into a live map of road conditions using computer vision and deep learning.

"Road Quality Assessment System Using Computer Vision and GPS-Based Mapping"

Anish Anilkumar  ·  2024TM93051  ·  M.Tech Software Engineering

Supervisor: Dr. Kavya Manohar  ·  Additional Examiner: Mr. Ashik Salahudeen  ·  BITS Evaluator: Dr. Eht E Sham

Explore the Maps Research Output ↓
SCROLL
Interactive Maps

EXPLORE THE DATA

Two AI models, two levels of granularity. Click any road segment to see observation counts, confidence scores, and raw label breakdowns.

5-Class Model

FULL SPECTRUM

Fine-grained classification across five conditions: Excellent → Good → Fair → Poor. Trained via active learning on thousands of manually-reviewed dashcam frames. Reveals the full gradient of road degradation.

Poor
Fair
Good
Excellent
Invalid
Open 5-Class Map
3-Class Model

BINARY SPLIT

High-confidence classification into Good and Bad roads. When the model is uncertain, it withholds judgement. Ideal for quick infrastructure triage — a clear red/green picture of what needs urgent attention.

Bad
Good
Invalid
Open 3-Class Map
Academic Research

RESEARCH OUTPUT

Reports and presentations from each stage of the M.Tech thesis, from initial outline through to the final submission.

Abstract / Outline
Project Outline Report
Initial research proposal defining the problem scope, objectives, and methodology for AI-powered road quality assessment.
Available Open PDF
Abstract / Outline
Outline Presentation
Slide deck presented for the abstract review stage, covering project motivation, pipeline design, and anticipated contributions.
Available Open PDF
Mid-Semester
Mid-Semester Report
Progress report covering the completed data pipeline, model architecture sweep across 15 backbones, and hyperparameter tuning results (72 trials).
Available Open PDF
Mid-Semester
Mid-Semester Presentation
Interactive HTML slide deck for the mid-semester review, with model comparison tables and live map demonstration.
Available Open Slides
Final
Final Report
Complete thesis document with full experimental results, ablation studies, and conclusions.
Coming Soon Not yet available
Final
Final Presentation
Thesis defence presentation summarising the end-to-end pipeline, model performance, and road quality mapping outcomes.
Coming Soon Not yet available
Research Artifact
Trained Classifier Model
Pre-trained Swin Transformer weights for both 5-class and 3-class road quality classification, ready for inference.
Coming Soon Not yet available
Research Artifact
Annotated Road Dataset
Crowd-sourced road condition classifications across thousands of dashcam frames, labelled by human annotators across 4 cameras.
Coming Soon Not yet available
Live Tool
Crowd-Source Annotation Platform
The collaborative, gamified road image labelling app built to generate training data for this project — open to contributors.
Open Source
Source Code
Full pipeline source — frame extraction, OCR, active learning classifier, YOLO pothole detector, and map renderer.
Methodology

HOW IT WORKS

A fully automated pipeline — from raw dashcam footage to a published, GPS-accurate road quality map.

01
Dashcam Video
Raw MP4
multi-contributor
02
Frame Extraction
1 fps via FFmpeg
hw accelerated
03
OCR Telemetry
GPS · speed · time
Tesseract OCR
04
AI Classification
Swin Transformer
active learning
05
GPS Snap
OpenStreetMap
segment snap
06
Interactive Map
Folium · colour-coded
road quality
Built with
Python PyTorch Swin Transformer YOLOv8 Tesseract OCR OpenStreetMap Folium osmnx FFmpeg