KSP Datathon 2024

Accident Severity ML

Python, CatBoost, Streamlit · May 2024

A machine learning web app that predicts motor vehicle accident severity (Fatal or Non-Fatal) and recommends emergency deployment levels using Karnataka State Police FIR data. Trained on 329,000+ real accident records spanning 38 districts across 2016–2024.

329k+
Accident Records
38
Districts
2016–24
Year Range
~59%
Test Accuracy
accident-analysis-ml.streamlit.app

Severity Prediction

Classifies accidents as Fatal or Non-Fatal from district, road type, month, year.

Deployment Level

Maps prediction to High Deployment or Standard Response with guidance.

Insights Dashboard

Five charts — monthly trends, severity split, top districts, road types, fatality rate by year.

Real KSP Data

329,000+ FIR records from 38 Karnataka districts, 2016–2024.

ML Model CatBoost — gradient boosting with native categorical support
Web App Streamlit with dark theme
Data Pipeline Pandas — 1.69M FIR rows filtered and deduplicated
Evaluation scikit-learn — stratified 80/20 split
Deployment Streamlit Community Cloud, auto-deploys from main

Built during KSP Datathon 2024 — a competition organised by Karnataka State Police to extract insights from their open FIR dataset. The dataset spans 1.69 million FIR records across all crime categories; we filtered to motor vehicle accidents and trained a classifier to distinguish fatal from non-fatal outcomes using contextual features available at report time.

The ~59% accuracy honestly reflects the constraint: fatal vs non-fatal outcome depends heavily on vehicle speed and collision type, which are not recorded in the FIR master table. The model is useful as a baseline risk signal from location and time context alone.

Shivam Sharma  ·  Krishnan Lakshmi Narayana  ·  Bhavya Vishal  ·  Madhuri