Inspiration
Every year, millions of lives are affected by road crashes — many of which could have been prevented with better data-driven insights. The idea for Zero Crash Project came from observing how most traffic systems are reactive — they respond after an accident happens.
We wanted to create a system that is proactive, one that can predict potential crashes before they occur. Inspired by smart city initiatives and the growing role of artificial intelligence in transportation, we envisioned a platform that combines computer vision, weather intelligence, and driver feedback to identify and reduce road accident risks.
Our guiding principle became clear: Predict. Prevent. Protect.
What It Does
Zero Crash Project is an AI-powered system that predicts the likelihood of road accidents by analyzing three main data sources:
CCTV Image Data – analyzed using a computer vision model trained on road and traffic images.
Weather Data – processed through a predictive model that considers rain, temperature, wind, and other environmental factors.
Driver Reports – text or voice input analyzed through Natural Language Processing (NLP) to classify road safety levels.
By integrating these three models, the system produces a crash risk level for each road segment, helping authorities and transportation departments take preventive actions in real time. R=f(V,W,D) Where: V = Vision Model output (image-based risk) W = Weather Model output (environmental risk) D = Driver Text Model output (human feedback risk) and f represents the fusion function that combines the three model outputs to produce the overall risk score R.
How We Built It
Computer Vision Model (Vertex AI Vision)
Trained with over 1,000 labeled road images covering different weather and lighting conditions.
Used Google Cloud Vertex AI for scalable model training and deployment.
Weather Prediction Model (Accuracy Boost)
Collected weather data such as temperature, rainfall, and wind speed.
Trained a regression model to correlate weather factors with crash probability.
Driver Text Model (Gemini API)
Integrated Google’s Gemini API for natural language text analysis.
Classifies driver comments as safe or risky based on content and tone.
Integration and Dashboard (Streamlit)
Combined all models into one interactive Streamlit dashboard.
Users can upload images, fetch real-time weather data, and input driver reports.
The final combined prediction and visualization are generated instantly.
Challenges We Ran Into
Data Scarcity: Collecting and labeling diverse CCTV datasets was time-consuming and limited.
Model Integration: Combining outputs from three independent AI models required synchronization and normalization of results.
Accuracy Optimization: The initial accuracy was moderate due to limited datasets and short training time.
Real-Time Data Issues: Integrating live APIs for weather and CCTV feeds caused delays and latency.
Text Data Quality: Driver reports often included informal language, abbreviations, or incomplete statements, making NLP classification difficult.
Each of these challenges taught us valuable lessons about scalability, data quality, and AI integration.
Accomplishments That We're Proud Of
Built and deployed three functional AI models covering vision, weather, and text analytics.
Created a fully integrated system that demonstrates the entire workflow end-to-end.
Achieved consistent crash-risk predictions during internal testing.
Developed a working prototype dashboard for visualization and interaction.
Proved that AI can be effectively applied to save lives by predicting road crashes before they occur.
What We Learned
The importance of multi-modal AI systems that combine image, numeric, and text data for better accuracy.
How to effectively use tools like Google Vertex AI, Gemini API, and Streamlit together in one unified project.
The mathematical relationships between weather conditions and accident probability.
That high-quality and diverse datasets directly lead to higher accuracy and reliability.
How interdisciplinary collaboration between data science, software engineering, and transport safety can create meaningful innovation.
What's Next for ZeroCrash
The future of Zero Crash Project focuses on scalability, automation, and real-time prediction.
Planned Improvements:
Integrate live CCTV camera feeds instead of static images.
Connect continuously updating weather APIs for real-time conditions.
Develop a mobile application for drivers to submit automatic voice or text feedback.
Implement a live monitoring dashboard for authorities to track accident-prone zones dynamically.
Improve model accuracy using larger datasets and advanced AI architectures such as YOLOv9 and transformer-based model fusion.
Our long-term vision is to help cities and governments move toward a future of zero road crashes, ensuring safer roads for everyone.
