Tech Stack
Python
Deep Learning
Computer Vision
PyTorch
TensorFlow
Description
Developed a comprehensive deep learning project implementing multiple semantic segmentation models for drone imagery analysis, focusing on safe landing spot detection for autonomous drones.
Implemented three distinct models: RGB model for standard image segmentation, RGBD model combining RGB and depth information, and MARS model specialized for planetary terrain analysis using advanced UNet++ architecture.
Achieved 85%+ IoU on validation set with the RGB model, enhanced segmentation accuracy through depth-aware features in RGBD model, and optimized MARS model for safe landing spot detection on Mars terrain with ONNX export for efficient deployment.
- RGB Model: Standard semantic segmentation achieving 85%+ IoU on drone imagery
- RGBD Model: Multi-modal fusion combining RGB and depth for enhanced accuracy
- MARS Model: UNet++ architecture with attention mechanisms for planetary analysis
- Advanced data augmentation using Albumentations for robust training
- Real-time inference capabilities with ONNX model export for deployment
- Safe landing spot detection algorithms with terrain risk assessment
- Comprehensive evaluation metrics and visualization tools