Anekha Sokhal

Portfolio showcasing my work in AI and Computer Vision

2 minute read

This is Part 1 of a two-part series on crater detection using deep learning. Read Part 2 →

Introduction 🚀

Lunar Navigation

As cislunar space grows increasingly crowded, spacecraft face heightened risks during navigation. Traditionally, space missions rely on Earth-based communication to determine their location. But what happens when communication is lost?

To solve this, NASA is developing TRON (Target & Range-adaptive Optical Navigation) — a next-gen system that allows spacecraft to determine their position independently, without Earth’s help.

TRON uses two methods:

  1. Planetary Triangulation for long-range navigation by measuring angles between celestial bodies.
  2. Crater-Based Localization for close-range navigation by detecting and matching craters seen in images with pre-mapped craters on the lunar surface.

This project focuses on crater detection — a key component of crater-based localization. Accurate crater recognition helps spacecraft understand their exact position, enabling autonomous operation in deep space.

📺 Watch a 1-minute project summary:

Project Objective 🌕

The goal is to build a system that automatically detects lunar craters in spacecraft images and fits an ellipse around each crater’s rim.

This is challenging due to:

  • Visual complexity: Craters vary in size, lighting, and visibility.
  • Hardware constraints: Must run on CPU (no GPU), under 20s per 5MP image, < 4GB RAM (like a Raspberry Pi 5).

These demand fast, lightweight deep learning models that can handle real-world noise and variability.

Literature Review 🧠

Early crater detection used edge detection and hand-crafted features, but struggled with shape and lighting variations.

Modern deep learning approaches:

  • Single-stage (e.g., YOLO, SSD) for speed.
  • Two-stage (e.g., Faster R-CNN, Mask R-CNN) for precision.

We compare:

  • YOLOv10: Fast, efficient real-time bounding box detector.
  • Ellipse R-CNN: Detects and fits elliptical shapes to crater rims.

Dataset 📸

We used a synthetic lunar crater dataset developed by NASA (via the Boromir engine), based on the crater catalog by Robbins (2019).

🗂️ Composition:

  • 1,000 grayscale images (2592×2048 px)
  • Paired JSON metadata: camera pose, sun angle, etc.
  • Dual label formats: YOLO (bounding boxes) & EllipseRCNN (elliptical params)

🖼️ Dataset Preview

Example Frames

⚠️ Challenges & Fixes

  • Misaligned masks from auto-generation
  • Extensive manual refinement using Label Studio
  • Ellipses converted to bounding boxes for YOLO

4. Experiments

🧪 Train/Test Split:

80/20 for both models

Evaluation Metrics

We use Intersection over Union (IoU) and Average Precision (AP) from PR curves to assess model accuracy and ranking.

IoU Metric

YOLOv10 📦

  • Model: YOLOv10m (mid-sized, balance of speed/accuracy)
  • Trained for 100 epochs using Ultralytics
  • Images resized to 640×640

Results:

  • Precision: 65%
  • Recall: 51%
  • mAP@0.5: 57%
  • COCO-style mAP (0.5–0.95): 34%

YOLO Results

Ellipse R-CNN 🌀

  • Backbone frozen, trained only on proposal, regression & occlusion modules
  • Tiling (512×512) significantly improved detection
  • Trained for 40 epochs, batch size 32

Results (on hold-out 100 images):

  • mAP@0.5: 17.2%
  • mAP@0.7: 9.9%

5. Next Steps 🧭

🔄 Evaluation Alignment

Define consistent dataset splits across YOLO and EllipseRCNN to ensure fair model comparison.

Pipeline Diagram

🔧 EllipseRCNN

  • Apply tiling during training
  • Explore optimizers, learning schedules
  • Address patch edge inference issues

⚙️ YOLOv10

  • Use full 1000-image dataset
  • Add tiling + augmentation
  • Train for 300 epochs
  • Try larger YOLOv10 variants
  • Apply same eval metrics as EllipseRCNN

You May Also Enjoy