Inspiration
About 16,800 police-reported crashes happen every day in the U.S. Afterward, everyone is trying to answer the same question: what actually happened?
A dashcam helps, but it only shows one angle. The car behind you might have recorded the part yours missed. That footage could be sitting on someone’s memory card without them knowing anyone needs it.
That’s the idea behind CrashDaddy: bring those perspectives together and make a crash easier to understand.
What it does
CrashDaddy helps you explore crash footage alongside an interactive 3D replay.
- Upload a dashcam video, along with any photos, notes, or supporting files.
- Explore reconstructed scenes, scrub through the timeline, and watch the original footage alongside them.
- Save recontructions and choose whether to share your footage.
- Search shared recordings by location and time to find other possible views of an incident.
- Our AI integration is designed to describe key moments, read visible license plates, estimate vehicle paths, suggest related recordings, and answer questions about the evidence.
How we built it
- Frontend: React and TypeScript
- Backend: FastAPI, PostgreSQL, and a Python worker for accounts, saved uploads, and background processing.
- Video processing: FFmpeg prepares clips and extracts frames for analysis.
- AI: Vision and language model adapters for scene analysis and investigation chat, plus a Roboflow integration for plate reading.
- Maps: A Python pipeline turns OpenStreetMap roads and buildings into 3D environments.
Challenges we ran into
- Deciding whether two recordings show the same crash. Being nearby at the same time isn’t enough, so the matching pipeline also checks for shared details.
- Keeping the video and 3D replay on the same clock as users pause, seek, and change playback speed.
- Separating what the footage actually shows from what the model estimates. A plausible-looking vehicle path still needs validation.
Accomplishments that we're proud of
- Watching the 3D replay move alongside the original footage and being able to pause and explore the scene.
- Connecting uploads, maps, and investigation tools into one experience.
- Building a way for someone’s dashcam footage to help another driver—not just themselves.
What we learned
“It renders” ≠ “it’s accurate.”
What's next for CrashDaddy
- Validate the AI pipeline on unfamiliar dashcam footage and measure how well the estimated paths and timing match.
- Align footage from different cameras in the same 3D scene.
- Test whether the replay actually helps people understand a crash faster and more accurately.
- Further develop physics to 3D recontructions
Log in or sign up for Devpost to join the conversation.