CrashWatch

Crashwatch uses a ML model to detect car crashes in real time from live traffic cameras and instantly alert first responders through automated voice reports, cutting response times and saving lives.

Comment

Inspiration

Last summer, one of us was in a car accident at 1:00 AM with nobody else around. If both drivers had been seriously hurt, they would have had no way to call for help. Hence, we created CrashWatch to ensure no one is ever left stranded after an accident by building a system that automatically reports crashes to get help on the way immediately.

What It Does

CrashWatch monitors live video from Maryland's public traffic cameras, as they are one of the few states that publishes the footage publically. Our custom-trained AI analyzes the footage in real-time, frame-by-frame, to detect accidents. A live dashboard displays each camera feed and shows the model's confidence level. When the system identifies a crash, it instantly generates an incident report and places an automated call to first responders.

How We Built It

We built CrashWatch as a complete, end-to-end system during the hackathon.

  • AI Model: We trained our own ResNet-34 model from scratch using crash footage from Kaggle.
  • Backend: A Python flask server powers the system, processing the live camera data, running the AI model, and serving the frontend.
  • Frontend: The monitoring page provides a real-time view of the camera feeds and any generated incident reports.
  • Automated Call System: The core of our system integrates three APIs. Twilio initiates the call, Google's Gemini API generates a clear script with the crash location and details, and ElevenLabs AI provides a natural, human-sounding voice to deliver the message.

Challenges We Ran Into

Training a computer vision model from scratch in a single weekend was our biggest challenge. We had to quickly process a large dataset and fine-tune the model for accurate predictions. Integrating three different APIs for the call system also proved complex, as we needed to orchestrate the data flow between them to create a seamless conversation. Finally, making the entire pipeline operate in real-time was an integration nightmare, where fixing one thing always broke something else.

Accomplishments We're Proud Of

Our proudest achievement is creating a fully automated system that solves a real-world problem. Training an effective AI model in such a short time was a major success. The automated call system, however, stands out the most. It’s more than a simple alert; it’s an intelligent agent that clearly communicates critical information to first responders, a feature that makes CrashWatch unique.

What We Learned

This project pushed us to build a practical AI application from concept to reality. We took a computer vision model from training to live deployment and learned to manage real-time video streams efficiently. By integrating Twilio, Gemini, and ElevenLabs, we mastered how to orchestrate multiple APIs into a single, intelligent workflow. The experience strengthened our skills in full-stack development, data processing, and applied AI.

What's Next for CrashWatch

We see CrashWatch as a vital tool for public safety. Next, we plan to expand its reach by adding camera networks from other states. We will also improve the AI's accuracy by training it on more diverse data, including different weather and lighting conditions. Finally, we want to enhance the automated call to provide even richer details, such as the number and types of vehicles involved.

Built With

Share this project:

Updates