Tracker Component Library (Python)

A Python port of the U.S. Naval Research Laboratory's Tracker Component Library, a comprehensive collection of algorithms for target tracking, estimation, coordinate systems, and related mathematical functions.

1,150+ functions | 190 modules | 6,700+ tests | 90% coverage

Overview

The Tracker Component Library provides building blocks for developing target tracking algorithms, including:

• Coordinate Systems: Conversions between Cartesian, spherical, geodetic, and other coordinate systems
• Dynamic Models: State transition matrices for constant velocity, coordinated turn, and other motion models
• Estimation Algorithms: Kalman filters (KF, EKF, UKF, CKF, CEKF, H-infinity), particle filters (bootstrap, RBPF), smoothers, and batch estimation
• Assignment Algorithms: Hungarian algorithm, auction algorithms, 3D/ND assignment, k-best assignments
• Data Association: Global Nearest Neighbor, JPDA, MHT for multi-target tracking
• Mathematical Functions: Special functions, statistics, numerical integration, and more
• Astronomical Code: SGP4/SDP4 propagation, TLE parsing, special orbits (parabolic/hyperbolic), ephemerides, relativistic corrections
• Reference Frames: GCRF, ITRF, TEME, TOD, MOD with full transformation chains
• Navigation: Geodetic calculations, INS mechanization, GNSS utilities, INS/GNSS integration
• Geophysical Models: Gravity (WGS84, EGM96/2008), magnetism (WMM, IGRF, EMM, WMMHR2025), atmosphere (US Standard 1976/ISA, simplified thermosphere, ionosphere), tides, terrain (GEBCO 2025, Earth2014)
• Signal Processing: Digital filters, matched filtering, CFAR detection, transforms (FFT, STFT, wavelets)
• GPU Acceleration: CuPy (NVIDIA CUDA) and MLX (Apple Silicon) backends for batch Kalman filtering and particle filters
• Results I/O: CSV/Parquet measurement readers, polars DataFrame accessors for track histories, msgspec (MessagePack/JSON) serialization, ASDF archival export, HDF5 compression (measured 4.73x), and AIS/NMEA transponder decoding
• Typed Configs & Sessions: msgspec.Struct configs (IMMConfig, GaussianSumConfig, RBPFConfig, SingleTargetConfig, MultiTargetConfig) accepted via a keyword-only config= on the matching constructor, plus full state snapshot/resume (pytcl.io.save_session/load_session) for six tracker and filter classes, bit-exact resume for the four deterministic ones

Installation

Basic Installation

pip install nrl-tracker

With Optional Dependencies

# For astronomy features (ephemerides, celestial mechanics)
pip install nrl-tracker[astronomy]

# For geodesy features (coordinate transforms, map projections)
pip install nrl-tracker[geodesy]

# For terrain data (GEBCO, Earth2014 via NetCDF)
pip install nrl-tracker[terrain]

# For visualization
pip install nrl-tracker[visualization]

# For signal processing (wavelets)
pip install nrl-tracker[signal]

# For polars DataFrame accessors on track histories and metrics
pip install nrl-tracker[dataframe]

# For AIS/NMEA transponder decoding
pip install nrl-tracker[ais]

# For ASDF archival export/import of tracks and states
pip install nrl-tracker[asdf]

# For GPU acceleration (NVIDIA CUDA)
pip install nrl-tracker[gpu]

# For GPU acceleration (Apple Silicon M1/M2/M3)
pip install nrl-tracker[gpu-apple]

# Install every user-facing extra except gpu (dev tooling is no longer a
# published extra — contributors use `uv sync`)
pip install nrl-tracker[all]

From Source

git clone https://github.com/nedonatelli/TCL.git
cd TCL
pip install -e .

Quick Start

Coordinate Conversions

import numpy as np
from pytcl.coordinate_systems import cart2sphere, sphere2cart

# Convert Cartesian to spherical coordinates
cart_point = np.array([1.0, 1.0, 1.0])
r, az, el = cart2sphere(cart_point, system_type="az-el")  # tracking convention
print(
    f"Range: {r:.3f}, Azimuth: {np.degrees(az):.1f}°, Elevation: {np.degrees(el):.1f}°"
)

# Convert back
cart_recovered = sphere2cart(r, az, el, system_type="az-el")

Kalman Filter

import numpy as np
from pytcl.dynamic_estimation.kalman import kf_predict, kf_update
from pytcl.dynamic_models import f_constant_velocity, q_constant_velocity

# Constant-velocity model, 2D state [x, vx, y, vy]
dt = 0.1
F = f_constant_velocity(dt, num_dims=2)
Q = q_constant_velocity(dt, sigma_a=1.0, num_dims=2)
H = np.array([[1.0, 0, 0, 0], [0, 0, 1.0, 0]])  # measure position only
R = np.eye(2) * 10.0

x = np.zeros(4)
P = np.eye(4) * 100.0
measurement = np.array([1.2, -0.7])

pred = kf_predict(x, P, F, Q)
upd = kf_update(pred.x, pred.P, measurement, H, R)
print(upd.x)  # updated state; upd.P, upd.y, upd.S, upd.K, upd.likelihood

Assignment Problem

import numpy as np
from pytcl.assignment_algorithms import hungarian

cost_matrix = np.array(
    [
        [10.0, 5.0, 13.0],
        [3.0, 15.0, 8.0],
        [7.0, 9.0, 12.0],
    ]
)

row_ind, col_ind, total_cost = hungarian(cost_matrix)
print(f"rows {row_ind} -> columns {col_ind}, total cost {total_cost}")  # cost 20.0

GPU Acceleration

The library supports GPU acceleration for batch processing of multiple tracks:

from pytcl.gpu import is_gpu_available, get_backend, to_gpu, to_cpu

# Check GPU availability (auto-detects CUDA or Apple Silicon)
if is_gpu_available():
    print(f"GPU available, using {get_backend()} backend")

    # Transfer data to GPU
    x_gpu = to_gpu(states)  # (n_tracks, state_dim)
    P_gpu = to_gpu(covariances)  # (n_tracks, state_dim, state_dim)

    # Use batch Kalman filter operations
    from pytcl.gpu import batch_kf_predict

    x_pred, P_pred = batch_kf_predict(x_gpu, P_gpu, F, Q)

    # Transfer results back to CPU
    x_pred_cpu = to_cpu(x_pred)

Supported backends:

• NVIDIA CUDA: Via CuPy (pip install nrl-tracker[gpu]) — float64
• Apple Silicon: Via MLX (pip install nrl-tracker[gpu-apple]) — float32

The backend is automatically selected based on your platform. Batch Kalman, EKF, UKF, particle-filter, and matrix operations all run on either backend.

Measured on Apple Silicon (MLX), batch linear Kalman predict+update versus a per-track CPU loop, end-to-end including host-device transfers and result materialization, after warm-up: 1.6x at 100 tracks, 13x at 1,000, 40x at 20,000 (August 2026).

Precision note: MLX computes in float32 (it raises on float64 GPU operations), so results match the CPU implementations to ~1e-7 relative rather than machine epsilon. The unscented filter is especially sensitive: its default alpha=1e-3 yields sigma-point weights of order 1e6, which float32 cannot resolve — use alpha >= 0.1 on MLX (the library warns).

Module Structure

pytcl/
├── core/                    # Foundation utilities and constants
├── mathematical_functions/  # Basic math, statistics, special functions
├── coordinate_systems/      # Coordinate conversions and transforms
├── dynamic_models/          # State transition and process noise models
├── dynamic_estimation/      # Kalman filters, particle filters
├── static_estimation/       # ML, least squares estimation
├── assignment_algorithms/   # 2D and multi-dimensional assignment
├── clustering/              # Mixture reduction, clustering
├── performance_evaluation/  # OSPA, track metrics
├── astronomical/            # Ephemerides, time systems
├── navigation/              # Geodetic, INS, GNSS
├── atmosphere/              # Atmosphere models, refraction
├── gravity/                 # Gravity models
├── magnetism/               # Magnetic field models
├── terrain/                 # Terrain elevation models
├── containers/              # Spatial indexes, track/measurement containers
├── trackers/                # Multi-target trackers (GNN, JPDA, MHT)
├── io/                      # SQL/HDF5 storage, CSV/Parquet/DataFrame/ASDF/msgspec I/O
├── transponders/            # AIS/NMEA decoding (maritime)
├── gpu/                     # GPU acceleration (CuPy/MLX)
└── plotting/                # Covariance ellipses, tracks, metrics plots

Examples & Tutorials

The library includes 42 runnable code examples demonstrating all major features:

Examples (32 files in /examples/)

Comprehensive demonstrations of library functionality:

• Tracking & Estimation: Kalman filters, particle filters, smoothers
• Assignment: Hungarian algorithm, k-best assignments, 3D assignment
• Coordinates: Frame conversions, transformations, geodetic calculations
• Dynamics: State models, motion models, dynamic systems
• Filtering: Uncertainty visualization, multi-target tracking
• Astronomy: Ephemerides, orbital mechanics, relativistic corrections
• Navigation: INS/GNSS integration, geophysical modeling
• Signal Processing: Detection, filtering, transforms
• Terrain & Atmosphere: Elevation models, atmospheric properties

Status: ✅ All 32 examples run in CI on every push

Tutorials (10 modules in /docs/tutorials/)

Interactive learning modules with visualizations:

• Assignment algorithms and 3D assignment problems
• Atmospheric and geophysical models
• Dynamical systems and reference frames
• Filtering and smoothing techniques
• Sensor fusion and advanced filtering
• Special functions and mathematical tools

Status: ✅ All 10 tutorials validated and passing (100% execution success)

Documentation

• API Reference
• User Guides
• Examples - 32 validated example scripts
• Tutorials - 10 interactive tutorial modules

Comparison with Original MATLAB Library

The core tracking workflow is fully ported and validated against independent references; the complete function-level accounting is in docs/matlab_parity_inventory.rst, and the explicit name mappings plus calling-convention differences are in docs/matlab_migration_map.rst. A taste:

Key differences:

• Function names use snake_case; multiple return values become NamedTuples
• States are 1-D arrays and batches are (N, dim) row-major (coordinate conversions also accept MATLAB-style column layouts)
• 0-based indexing, explicit unassigned_rows instead of 0-sentinels

Testing

# Run all tests
uv run pytest

# Run with coverage
uv run pytest --cov=pytcl

# Run only fast tests
uv run pytest -m "not slow"

# Run tests validated against MATLAB
uv run pytest -m matlab_validated

A bare local pytest run picks up tests/property/'s Hypothesis-generated property tests too, at the 500-example dev profile (CI runs a derandomized 100-example profile instead); see tests/property/README.md for what's covered and the profile policy.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

Development Setup

git clone https://github.com/nedonatelli/TCL.git
cd TCL
uv sync
uv run prek install

See CONTRIBUTING.md for the full setup (installing uv, uv run vs. activating .venv, the gpu-apple extra for Apple Silicon).

Running Quality Checks

# Format code
uv run ruff format .

# Lint (includes import sorting)
uv run ruff check .

# Type check (gate)
uv run ty check pytcl

# Run all checks
uv run prek run --all-files

Citation

If you use this library in your research, please cite the original MATLAB library:

@article{crouse2017tracker,
  title={The Tracker Component Library: Free Routines for Rapid Prototyping},
  author={Crouse, David F.},
  journal={IEEE Aerospace and Electronic Systems Magazine},
  volume={32},
  number={5},
  pages={18--27},
  year={2017},
  publisher={IEEE}
}

License

This project is in the public domain, following the original MATLAB library's license. See LICENSE for details.

Acknowledgments

• Original MATLAB library by David F. Crouse at the U.S. Naval Research Laboratory
• This port follows the Federal Source Code Policy (OMB M-16-21)

Related Projects

• FilterPy - Kalman filtering library
• Stone Soup - Framework for tracking algorithms
• Astropy - Astronomy library for Python