Python GLMB filter with History-based Birth

This code was used to produce the results shown in:

C. Aguilar, M. Ortner and J. Zerubia, "Adaptive Birth for the GLMB Filter for object tracking in satellite videos," 2022 IEEE 32st International Workshop on Machine Learning for Signal Processing (MLSP), 2022, pp. 1-6, doi: "

If you use this code, we strongly suggest you cite:

@inproceedings{aguilar2022,
    author = {Aguilar, Camilo and Ortner, Mathias and Zerubia, Josiane},
    title = {Adaptive Birth for the GLMB Filter for object tracking in satellite videos},
    booktitle = {2022 IEEE 32nd International Workshop on Machine Learning for Signal Processing (MLSP)},
    pages={1-6},
    doi={},
    Year = {2022}
}

Tracked Objects	Adaptive Birth Field

Installation

Clone the repository

git clone --recursive https://github.com/Ayana-Inria/GLMB-adaptive-birth-satellite-videos

To install required dependencies run:

$ conda env create -f ab_glmb.yml

Required libraries

- numpy=1.22
- pytorch=1.12
- opencv=4.5
- pillow=9.0
- matplotlib=3.5
- motmetrics=1.2

Usage

To run the demo

python demo.py

Demo outputs are saved under:

[root directory]/dataset/WPAFB_2009/AOI_02/FILTER_OUTPUT/

Parameter Choice

The GLMB filter relies on numerous parameters, but the most important ones are:

# parameters.py
parameters['tau'] = 1   # Sampling period
parameters['Q'] = 10    # Motion covariance  
parameters['R'] = 1     # Measurement covariance
parameters['Po'] = 5    # Birth covariance

Using AB-GLMB in your own project

All the essential functions for our filter are located at the src/filter/glmb.py file:

# python 3.6
import src.filter.glmb as GLMB_FILTER

# Fix filter parameters
parameters = {}
parameters['tau'] = 1   # Sampling period
parameters['Q'] = 10    # Motion covariance  
parameters['R'] = 1     # Measurement covariance
parameters['Po'] = 5    # Birth covariance

# Initialize Filter Model
model = GLMB_FILTER.Model(parameters)

# Create instance of GLMB
glmb_update = GLMB_FILTER.glmb_instance()

# Create birth field
birth_field = torch.zeros((data_rows, data_cols, 3))

# Create state list
X = []

# Iterate over each frame
    # get frame detections
    zk = get_frame_detections(frame_number) # torch tensor with shape (n_objects_at_frame_k, n_dims). n_dims=4: [px, py, w, h]

    # Update GLMB
    glmb_update = GLMB_FILTER.jointpredictupdate_a_birth(glmb_update, model,  zk, birth_field, frame_number)

    # State Estimation
    Xk = glmb_update.extract_estimates() # Xk is a dictionary of the form Xk[obj_label] = [px, py, vx, vy, w, h]

    # Add frame state to state list
    X.append(Xk)

    # Update birth field
    birth_field = update_birth_field(birth_field, X, frame_number)

Dataset

Video Regisration

Download the WPAFB 2009 dataset from the AFRL's Sensor Data Management System (SDMS) website
Convert the .ntf files to .png (we used Matlab's nitfread function)
Use our video stabilization repository to stabilize the sequence and format the labeling

Data Structure

To replicate the results shown in the paper, the DATASET needs to be formatted in the following way:

[root for demo.py]
└──dataset
      └── WPAFB_2009/
            └── AOI_02/
                ├── INPUT_DATA/
                |   ├── img01.png
                |   ├── img02.png
                |   └── ...
                ├── GT/
                |   ├── stabilized_oject_states.csv
                |   └── labels/
                |         ├── labels_as_points_01.png
                |         ├── labels_as_points_02.png
                |         └── ...
                └── FILTER_OUTPUT/
                    ├── birth_field/
                    |     ├── birth_field_01.png
                    |     ├── birth_field_02.png
                    |     └── ...
                    ├── objects/
                    |     ├── tracked_objects_01.png
                    |     ├── tracked_objects_02.png
                    |     └── ...
                    ├── labels/
                    |     ├── labels_as_points_01.png
                    |     ├── labels_as_points_02.png
                    |     └── ...
                    └── object_states.csv

Aknowledgment

Thanks to BPI France (LiChiE contract) for funding this research work, and to the OPAL infrastructure from Université Côte d'Azur for providing computational resources and support.

License

GLMB with adaptive birth is released under the GNUv3 License (refer to the LICENSE file for details).

TaiPhillips/GLMB-adaptive-birth-satellite-videos