Skip to content

wdoppenberg/crater-detection

Repository files navigation

Lunar Orbit Navigation Using Ellipse R-CNN and Crater Pattern Matching

Crater Detection Sample

This repository contains the necessary components to compose a Terrain-Relative Navigation (TRN) pipeline for navigating around the moon at altitudes below 500km. It utilizes a crater rim detection method based on Faster R-CNN & Ellipse R-CNN which is then trained to perform its task on simulated monocular camera images. These images are generated using SurRender, and the training mask is created by projecting the apparent ellipses of craters into the image plane. Identification is done using the method as described in Christian et al., 2020, which allows the system to create a database of projective invariants that are unique to each (approximately) coplanar crater triad.

Note: The notebooks in this repository are for development purposes and contain deprecated code - they will be updated in the future to demonstrate the capabilities of the package.

Model description

Crater detector

Instructions

This repository uses git LFS.

Dataset creation (SurRender server required)

usage: create_dataset.py [-h] [--n_train N_TRAIN] [--n_val N_VAL]
                         [--n_test N_TEST] [--identifier IDENTIFIER]
                         [--resolution RESOLUTION] [--fov FOV]
                         [--min_sol_incidence MIN_SOL_INCIDENCE]
                         [--max_sol_incidence MAX_SOL_INCIDENCE]
                         [--ellipse_limit ELLIPSE_LIMIT] [--filled FILLED]
                         [--mask_thickness MASK_THICKNESS]

Create dataset containing input images and relevant ellipses using SurRender.

optional arguments:
  -h, --help            show this help message and exit
  --n_train N_TRAIN     Number of training images (default: 20000)
  --n_val N_VAL         Number of validation images (default: 1000)
  --n_test N_TEST       Number of testing images (default: 1000)
  --identifier IDENTIFIER
                        Number of testing images (default: None)
  --resolution RESOLUTION
                        Camera resolution (default: (256, 256))
  --fov FOV             Camera FoV (default: 45)
  --min_sol_incidence MIN_SOL_INCIDENCE
                        Minimum solar incidence angle (default: 10)
  --max_sol_incidence MAX_SOL_INCIDENCE
                        Maximum solar incidence angle (default: 80)
  --ellipse_limit ELLIPSE_LIMIT
                        Maximum ellipticity for gt ellipse shapes
                        (selenographic) (default: 1.3)
  --filled FILLED       Whether to fill the crater masks or not (default:
                        True)
  --mask_thickness MASK_THICKNESS
                        How thick to make the mask rim if not filled (default:

Model training

Start MLflow server using:

mlflow server --backend-store-uri=sqlite:///mlruns.db --default-artifact-root=artifacts

After installing all required packages and retrieving the dataset, start training using train_model.py:

usage: train_model.py [-h] [--epochs EPOCHS] [--batch_size BATCH_SIZE]
                      [--num_workers NUM_WORKERS]
                      [--learning_rate [LEARNING_RATE]] [--backbone BACKBONE]
                      [--run_id [RUN_ID]] [--dataset DATASET]
                      [--momentum MOMENTUM] [--weight_decay WEIGHT_DECAY]
                      [--device DEVICE]

Train the crater detection model on images and target ellipses

optional arguments:
  -h, --help            show this help message and exit
  --epochs EPOCHS       Number of epochs (default: 20)
  --batch_size BATCH_SIZE
                        Batch size (default: 8)
  --num_workers NUM_WORKERS
                        Number of workers for training dataloader (default: 4)
  --learning_rate [LEARNING_RATE]
                        Learning rate (default: 0.001)
  --backbone BACKBONE   Model backbone ResNet type. (default: resnet50)
  --run_id [RUN_ID]     Resume from MLflow run checkpoint (default: None)
  --dataset DATASET     Dataset path (default:
                        data/dataset_crater_detection_40k.h5)
  --momentum MOMENTUM   Momentum input for SGD optimizer. (default: 0.9)
  --weight_decay WEIGHT_DECAY
                        Weight decay input for SGD optimizer. (default: 1e-07)
  --device DEVICE       Device to train model on (`cpu` or `cuda`) (default:
                        cuda)