thresh

Multi-sensor fusion multi-object tracking framework in Rust.

Hybrid architecture: transformer-based detection (via ONNX Runtime) + classical Bayesian state estimation (Kalman filter family). Designed for heterogeneous aerospace targets spanning UAVs through ballistic missiles.

Workspace

Crate	Description
thresh-core	Common types: state vectors, measurements, covariance matrices, coordinates
thresh-filter	KF, EKF, UKF with CV, CA, CTRV, Coordinated Turn motion models
thresh-association	Hungarian algorithm, Mahalanobis gating, 2D/3D IoU, cascaded association
thresh-fusion	Centralized fusion, information filter, covariance intersection
thresh-inference	ONNX Runtime pipeline orchestration (feature-gated)
thresh-tracker	Track lifecycle, M-of-N confirmation, class-specific heads
thresh-bridge	PyO3 bridge to Stone Soup (feature-gated)
thresh-synth	Synthetic trajectory + sensor data generation
thresh-eval	MOT metrics: MOTA, MOTP, IDF1, HOTA, AMOTA

Quick start

use thresh_tracker::tracker::MultiObjectTracker;
use nalgebra::DVector;

// Create a tracker with 10m measurement noise, 100 chi-squared gate
let mut tracker = MultiObjectTracker::new_cv_position(10.0, 100.0);

// Feed detections each frame
let detections = vec![
    DVector::from_column_slice(&[1000.0, 2000.0, 5000.0]),
];
tracker.step(&detections, 1.0); // dt = 1 second

Building

cargo build --workspace
cargo test --workspace

Optional features:

• onnx on thresh-inference: enables ONNX Runtime (requires runtime binaries)
• stonesoup on thresh-bridge: enables PyO3 Stone Soup integration
• jsbsim, rcs-compute, radar-scene on thresh-synth: PyO3 bridges to JSBSim, PyPOFacets, and RadarSimPy for high-fidelity sensor simulation (all require a Python environment with the corresponding package installed)
• adsb, orbital, nuscenes on thresh-data: dataset adapters (HTTP, SGP4, PyO3)

Sensor fidelity levels

thresh-synth supports three progressive fidelity levels for radar measurement generation. Pick the lowest that still answers the question you're asking; each step up adds realism but also randomness, runtime cost, and/or external dependencies.

Level	Module / entry point	Detection model	RCS model	When to use
0 — Simple	measurement_gen::generate_radar with RadarConfig { radar_equation: None, p_detection: 0.85, … }	Fixed p_detection	Not modeled	Algorithm development, deterministic regression tests, tracker integration tests where you want to isolate tracker behaviour from sensor noise.
1 — Radar equation	radar_equation::generate_radar_full with FullRadarConfig { radar: RadarParameters::x_band_surveillance(), apply_atmosphere: true, use_shnidman: true, … } + swerling::RcsProfile::fighter() (or similar)	albersheim_pd / shnidman_pd from range-dependent SNR, ITU-R P.676 atmospheric loss, Swerling I/II/III/IV fluctuation	Swerling chi-squared distribution + optional aspect-angle lookup table (RcsLookupTable)	Realistic P_d vs range behaviour, tracker robustness testing under realistic drop-outs, benchmarks that need physically meaningful baselines. Pure Rust — runs in default CI.
2 — Full wave/ray simulation	radar_scene::RadarSceneBridge (feature radar-scene) → PyO3 to RadarSimPy; rcs_compute::RcsComputeBridge (feature rcs-compute) → PyO3 to PyPOFacets	High-fidelity waveform / return-pulse simulation via RadarSimPy	Method-of-Moments / Physical Optics RCS computed from an STL geometry via PyPOFacets	Sensor design validation, proof that a specific radar / target / environment combination meets a requirement, data generation for transformer-based detectors. Requires a Python environment and is excluded from default CI builds because of the heavy external dependency.

The thresh-rcs-compute CLI binary (cargo install --path crates/thresh-synth --features rcs-compute) wraps Level 2 RCS computation so you can precompute an RcsLookupTable JSON from an STL file and then feed it into Level 1 without paying the PyO3 cost on every measurement.

End-to-end integration tests for each scenario live in crates/thresh/tests/hifi_integration.rs:

• hifi_radar_swerling_scenario — fighter trajectory, Level 1 radar with Swerling I fluctuation.
• hifi_orbital_radar_scenario — ISS-like overhead pass against a ground station radar, range-dependent P_d.
• hifi_multisensor_radar_eoir — radar + MWIR EO/IR physics fusion on a maneuvering aircraft.
• fidelity_level_comparison — same trajectory run at Level 0 and Level 1, logs the MOTA delta.

Run them with cargo test -p thresh --test hifi_integration.

License

Apache-2.0