A Deep Learning pipeline that perceives complex terrains and plans optimal trajectories in 20x20 grid environments.
This project tackles the challenge of autonomous navigation in unstructured environments. Given a raw map image containing various terrains (forests, deserts, labs), the system must:
- Perceive: Identify walkable paths, obstacles (walls), hazards, start points, and goal points using Computer Vision.
- Plan: Compute the safest and most efficient path from Start to Goal.
The solution integrates Deep Semantic Segmentation (ResNet-UNet) with Classical Graph Search (A)* to achieve near-perfect navigation performance.
- 🧠 Advanced Perception: Uses a ResNet34-UNet architecture to segment 20x20 maps with high pixel-wise accuracy.
- 🏭 Synthetic Data Engine: Includes a custom generator that creates thousands of synthetic training maps with realistic textures to overcome data scarcity.
- ⚡ Velocity-Aware A*: A custom A* implementation that handles variable terrain costs (avoiding hazards vs walls).
- 🛡️ Robust Inference: Implements Test-Time Augmentation (TTA) (flips/rotations) to ensure stable predictions on unseen data.
- 🎯 High Accuracy: Achieved >99% pathfinding success rate on the validation set.
| Domain | Technology |
|---|---|
| Deep Learning | PyTorch, Torchvision (ResNet34 Backbone) |
| Image Processing | OpenCV, PIL, Albumentations |
| Data Handling | NumPy, Pandas |
| Algorithms | A* (A-Star), BFS, Manhattan Geometry |
| Visualization | Matplotlib, Seaborn |