Computer Vision / PyTorch / Gradio
Terrain Safety Analysis & Navigation Assistant
Vision prototype for autonomous navigation using segmentation, depth, and a safety-score engine.
Problem
Assess traversability and visualize safe navigation paths for autonomous terrain traversal.
Approach
Used SegFormer for 150-class semantic segmentation and Depth Anything V2 for monocular depth. Built a safety-score algorithm with A* pathfinding to render optimal routes on a 3D terrain mesh. Added slope + morphology refinements to reduce false positives and wired W&B for experiment tracking.
Results
- 3D terrain mesh routing visualization
- Robust hazard detection with geometric refinement