blind-nav-ai

Arabic-speaking AI assistant for blind navigation using zero-shot vision-language models.

Repository Structure & Branches

• main — Stable integration point; holds scaffold and merged features.

• ml-quick — Quick track with real-time, low-latency models:

  • PyDNet for depth estimation (~60 FPS)
  • YOLOv8-Nano for object detection
  • BLIP-2 (tiny) for on-device captioning

• ml-best-base — Best track with high-accuracy models:

  • GroundingDINO for open-vocabulary grounding
  • MiDaS-small for depth estimation (~20–30 FPS)
  • BLIP-2 (tiny) for rich scene descriptions

Usage Modes & Expected Output

🔹 Quick Mode (ml-quick branch)

Pipeline: PyDNet → YOLOv8-Nano → BLIP-2 (tiny) → Template-based TTS

Visual Overlay:

• Bounding boxes around detected objects
• Distance labels (e.g., “2.1 m”) beside each box

Spoken Alerts: Succinct, templated sentences (up to 5 Hz). Example:

[Beep] “Chair one point five meters ahead.”
[Beep] “Person two point three meters ahead.”

🔹 Best-Base Mode (ml-best-base branch)

Pipeline: GroundingDINO → MiDaS-small → BLIP-2 (tiny) → Free-form planning → TTS

Visual Overlay:

• Precise region labels (open-vocabulary)
• Distance heatmap or per-object distance tags

Spoken Narration: Rich, free-form descriptions. Example:

[Beep] “There’s a wooden chair one point five meters in front of you, on your left.”
[Beep] “A person is walking about two point three meters ahead.”

Getting Started s

1. Clone the repo and checkout your track:

   git clone https://github.com/MohammadYusif/blind-nav-ai.git
   # Quick mode
   git checkout ml-quick
   # Best-base mode
   git checkout ml-best-base

2. Follow the branch-specific README in /ml-quick or /ml-best-base for setup and demo instructions.

---