ASCEND Vision System

ASCEND Vision System is the computer vision and perception stack developed for the ISRO Robotics Challenge (IRoC-U). The project focuses on autonomous terrain feature detection, image retrieval, visual localization, mapping, and navigation in GPS-denied environments using a hybrid approach that combines classical computer vision, deep learning embeddings, and SLAM.

The system leverages ORB feature extraction, MobileNetV3-based image embeddings, similarity search, and future ORB-SLAM3 integration to enable reliable feature recognition and autonomous decision-making on resource-constrained edge hardware such as the NVIDIA Jetson Orin Nano.

Features Implemented

Preprocessing

• Blur Detection

Classical Computer Vision

• ORB Feature Extraction
• ORB Feature Matching
• Lowe Ratio Test

Deep Learning

• MobileNetV3 Feature Extraction
• Embedding Generation
• Cosine Similarity Matching
• Image Retrieval System

Tech Stack

• Python
• OpenCV
• PyTorch
• TorchVision
• NumPy

Current Progress

• ORB Feature Pipeline
• MobileNet Embeddings
• Image Retrieval
• Hybrid Retrieval + ORB Verification
• ORB-SLAM3 Integration
• ROS2 Integration
• Jetson Deployment

Project Structure

ASCEND-Vision-System ├── preprocessing ├── matching ├── deep_learning ├── dataset ├── results