autonomous_drone_nav_vision
Overview
A Vision Transformer (ViT) fine-tuned for tactical aerial navigation. This model enables Small Unmanned Aircraft Systems (sUAS) to classify environmental obstacles and identify safe landing zones in real-time using downward and forward-facing RGB cameras.
Model Architecture
The model utilizes a Vision Transformer (ViT-Base) backbone:
- Patch Extraction: Images are divided into $16 \times 16$ fixed-size patches.
- Position Embeddings: Learnable spatial embeddings are added to the patch sequence to retain structural context.
- Attention Mechanism: Global self-attention allows the model to correlate distant visual cues, such as horizon lines and ground markers.
Intended Use
- Obstacle Avoidance: Integrated into flight control stacks for autonomous "sense and avoid" maneuvers.
- Precision Landing: Identifying designated markers or flat terrain for autonomous recovery.
- Search and Rescue: Preliminary screening of aerial footage to identify human-made structures or anomalies.
Limitations
- Low Light: Performance degrades significantly in nighttime or heavy fog conditions without thermal input.
- Motion Blur: Rapid yaw movements at high speeds may cause misclassification due to pixel streaking.
- Scale Invariance: Small objects at extreme altitudes may be missed due to the fixed $224 \times 224$ input resolution.
