Drive smarter. Drive greener. — SF Hacks 2026

EcoRoad AI analyzes dashcam (or any road) video with computer vision and gives you eco and safety insights — no car sensors or GPS required. Upload a clip, get an annotated video plus perception counts, scene type, risk level, eco score, CO₂ and fuel estimates, driving style, and actionable tips.

Demo video: Watch on YouTube

• Upload a short video (dashcam, phone, or any road clip; MP4, MOV, WebM).
• AI runs object detection (YOLO) and scene analysis: vehicles, pedestrians, traffic lights, signs, optical-flow speed estimate.
• You get:
  • Annotated video with detection boxes and a timeline of alert moments
  • Perception — counts (cars, people, lights, signs)
  • Decision — scene type, risk level, eco score (0–100), CO₂ g/km (or g/mi)
  • Recommendations — tips to drive smoother and use less fuel
  • Trip analytics — driving style, fuel & cost (L/gal, km/mi), CO₂ by phase, action plan, key moments, trends, savings what‑if, trip summary

Use the km / mi toggle in the nav to switch metric and imperial everywhere (fuel, distance, CO₂). Use Clear uploads & results to delete stored videos and outputs. Gallery shows all detection videos in one view.

Requirements: Python 3.10+, pip. GPU optional (faster inference with CUDA).

cd ecoRoadAI
python -m venv .venv

Windows (PowerShell):

.venv\Scripts\activate
pip install -r requirements.txt
python app.py

Mac / Linux:

.venv/bin/pip install -r requirements.txt
./run.sh

Then open http://localhost:5050. Default port is 5050 (to avoid conflict with macOS AirPlay on 5000). To use another port:

• Windows: set PORT=8080 && python app.py
• Mac/Linux: PORT=8080 ./run.sh

Upload a video and wait for analysis (~30–60 s depending on length and device). Click ⓘ icons for short explanations. Download video and Download results (JSON) save the annotated clip and full run data.

Try with sample: Put a short video at samples/sample.mp4 (or use clips in bdd100k/) and a Try with sample video button appears so you can run without uploading.

The app uses a pre-trained YOLO model (yolov8n.pt). Optional fine-tuning on BDD100K: see scripts/train_bdd100k.py.

Optional env:

• ECOROAD_MODEL_PATH — path to weights (default: yolov8n.pt)
• ECOROAD_CONF — detection confidence 0–1 (default: 0.25)
• ECOROAD_DEVICE — cuda or cpu
• ECOROAD_SAVE_FLOW_VIS — set to 1 to save a Farneback optical-flow visualization video (GPU/batched runs only). In the app, use the Detection | Optical flow toggle to switch the player to the flow view.

• README (this file) — overview, run instructions, tech, project layout.
• docs/ — Setup, Pipeline overview.

• Distance and speed are estimated from the video (scene + optical flow), not GPS. Short clips (e.g. 40 s) often show ~0.3–0.5 mi.
• Trip history and trends are stored under videos/analytics/ and persist until you click Clear uploads & results.

MIT License — see LICENSE.