norfair-rs

Real-time multi-object tracking for Rust

---

Disclaimer: This is an unofficial Rust port of Python's norfair object tracking library. This project is NOT affiliated with, endorsed by, or associated with Tryolabs or the original norfair development team. All credit for the original design and algorithms goes to the original norfair authors.

---

Overview

norfair-rs is a Rust implementation of the norfair multi-object tracking library, bringing real-time object tracking capabilities to Rust applications with:

• Detector-agnostic design: Works with any object detector (YOLO, Faster R-CNN, custom models)
• Rust-native performance: Zero-cost abstractions, no GC, maximum speed
• Type-safe API: Compile-time validation of tracking configurations
• Comprehensive Tests: 289+ Rust tests + 98 Python tests ensuring correctness and equivalence with the original norfair library
• Drop-In-Replacement: Python bindings with uv add norfair_rs and import norfair_rs as norfair

Related Projects

• norfair - Original Python implementation by Tryolabs
• norfair-go - Go port of norfair (sibling project)

Features

• Flexible Distance Functions: IoU, Euclidean, Manhattan, Frobenius, Keypoint Voting, and more
• Multiple Filtering Options: Optimized Kalman filter, full filterpy-equivalent Kalman, or no filtering
• Re-identification (ReID): Full support for object re-identification with reid_hit_counter lifecycle
• Thread-safe: Concurrent-safe ID generation and tracking
• Python Callable Support: Use custom Python distance functions directly with the Tracker

Benchmarks

Cross-language performance comparison (IoU distance, OptimizedKalmanFilter):

Scenario	Frames	Detections	norfair	norfair-go	norfair-rs (python)	norfair-rs (rust)
Small	100	446	4,700 fps	243,000 fps	107,000 fps	296,000 fps
Medium	500	9,015	540 fps	31,000 fps	27,000 fps	89,000 fps
Large	1,000	44,996	101 fps	3,800 fps	11,000 fps	41,000 fps
Stress	2,000	179,789	—	547 fps	5,200 fps	18,500 fps

Speedup norfair-rs (rust) vs norfair: 60-180x depending on scenario complexity Speedup norfair-rs (python) vs norfair: 20-50x (drop-in replacement)

Benchmarks run on Apple M3 Pro. See examples/benchmark/ for reproduction scripts.

Comparison

Play around with the norfair (python) vs norfair-rs (rust) comparison tool using:

uv run examples/compare_norfair_py_rs.py

---

Installation

Rust

Add to your Cargo.toml:

[dependencies]
norfair-rs = "0.3"

Python (Drop-in Replacement)

norfair-rs provides Python bindings that work as a drop-in replacement for the original norfair library, with 20-50x better performance:

uv add norfair-rs
# or: pip install norfair-rs

Then simply change your import:

# Before (original norfair)
from norfair import Detection, Tracker

# After (norfair-rs - same API, much faster!)
from norfair_rs import Detection, Tracker

Most of your existing norfair code should work unchanged. See the benchmark results for performance comparisons.

Python API Compatibility

Compatible with norfair:

• ✅ Detection, TrackedObject, Tracker - Same API
• ✅ Custom Python callable distance functions
• ✅ Custom Python callable reid_distance_function
• ✅ create_keypoints_voting_distance(), create_normalized_mean_euclidean_distance()
• ✅ All built-in distance functions via get_distance_by_name()
• ✅ Full ReID support with reid_hit_counter lifecycle
• ✅ Filter factories: OptimizedKalmanFilterFactory, FilterPyKalmanFilterFactory, NoFilterFactory
• ✅ TranslationTransformation for camera motion

Not yet implemented: The following features from Python norfair are not available in norfair-rs:

• Video I/O - Video class (requires OpenCV)
• Drawing - draw_boxes(), draw_points(), draw_tracked_objects(), Drawer, Paths (requires OpenCV)
• Homography - HomographyTransformation, MotionEstimator (requires OpenCV)
• Utilities - FixedCamera, get_cutout(), print_objects_as_table()
• Scipy - Scipy vectorised distance functions and some other features.

Most of these features require OpenCV bindings which are not yet implemented. Core tracking works fully.

Quick Start (Rust)

use norfair_rs::{Detection, Tracker, TrackerConfig};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // 1. Create tracker with IoU distance function
    let mut config = TrackerConfig::from_distance_name("iou", 0.5);
    config.hit_counter_max = 30;
    config.initialization_delay = 3;

    let mut tracker = Tracker::new(config)?;

    // 2. For each frame
    for frame in iter_video_frames() {
        // 2.1 Generate detections from your object detector
        let detections: Vec<Detection> = detect_objects(&frame)
            .iter()
            .map(|bbox| {
                // Bounding box format: [x1, y1, x2, y2]
                Detection::from_slice(
                    &[bbox.x, bbox.y, bbox.x + bbox.w, bbox.y + bbox.h],
                    1, 4  // 1 row, 4 columns
                ).unwrap()
            })
            .collect();

        // 2.2 Update tracker, returning current tracked objects with stable IDs
        let tracked_objects = tracker.update(detections, 1, None);

        // 2.3 Use tracked objects (draw, analyze, etc.)
        for obj in tracked_objects {
            if let Some(id) = obj.id {
                draw_box(&frame, &obj.estimate, id);
            }
        }
    }

    Ok(())
}

Python Norfair Equivalent

Here's how the same tracking workflow looks in the original Python norfair library:

Python:

# OLD: from norfair import Detection, Tracker
from norfair_rs import Detection, Tracker

# Create tracker
tracker = Tracker(
    distance_function="iou",
    distance_threshold=0.5,
    hit_counter_max=30,
    initialization_delay=3,
)

# Process frames
for frame in iter_video_frames():
    # Get detections from your detector
    detections = [
        Detection(points=np.array([[x1, y1, x2, y2]]))
        for x1, y1, x2, y2 in detect_objects(frame)
    ]

    # Update tracker
    tracked_objects = tracker.update(detections=detections)

    # Use tracked objects
    for obj in tracked_objects:
        draw_box(frame, obj.estimate, obj.id)

Key Differences:

• Rust: Explicit configuration structs vs Python kwargs
• Rust: Error handling with Result<T, E> returns
• Rust: Uses nalgebra matrices instead of numpy arrays
• Rust: Zero-cost abstractions with compile-time guarantees

Both implementations provide the same core functionality with Rust offering better performance.

Configuration Options

use norfair_rs::{TrackerConfig, filter::OptimizedKalmanFilterFactory};
use norfair_rs::distances::distance_by_name;

let mut config = TrackerConfig::new(distance_by_name("euclidean"), 50.0);

// Tracking behavior
config.hit_counter_max = 15;           // Frames to keep tracking without detection
config.initialization_delay = 3;       // Detections required to initialize
config.pointwise_hit_counter_max = 4;  // Per-point tracking threshold
config.detection_threshold = 0.5;      // Minimum detection confidence
config.past_detections_length = 4;     // History for re-identification

// Re-identification (optional)
config.reid_distance_function = Some(distance_by_name("euclidean"));
config.reid_distance_threshold = 100.0;
config.reid_hit_counter_max = Some(50);

// Kalman filter
config.filter_factory = Box::new(OptimizedKalmanFilterFactory::new(
    4.0,   // R (measurement noise)
    0.1,   // Q (process noise)
    10.0,  // P (initial covariance)
    0.0,   // pos_variance
    1.0,   // vel_variance
));

Distance Functions

Built-in distance functions available via distance_by_name():

Name	Description	Use Case
"euclidean"	L2 distance between points	Single-point tracking
"iou"	1 - Intersection over Union	Bounding box tracking
"mean_euclidean"	Average L2 across all points	Multi-keypoint tracking
"mean_manhattan"	Average L1 across all points	Grid-aligned tracking
"frobenius"	Frobenius norm of difference	Matrix comparison

Custom distance functions can be implemented via the Distance trait.

Filter Options

Three filter types are available:

use norfair_rs::filter::{
    OptimizedKalmanFilterFactory,  // Fast, simplified Kalman (default)
    FilterPyKalmanFilterFactory,    // Full filterpy-compatible Kalman
    NoFilterFactory,                // No prediction (detection-only)
};

API Documentation

Core Types

• Tracker - Main tracking engine that maintains object identities across frames
• Detection - Input from object detector (bounding boxes, keypoints, or arbitrary points)
• TrackedObject - Output object with stable ID, position estimate, and tracking metadata
• TrackerConfig - Configuration for tracker behavior
• TrackedObjectFactory - Thread-safe ID generation

Camera Motion

use norfair_rs::camera_motion::TranslationTransformation;

// Compensate for camera movement
let transform = TranslationTransformation::new([dx, dy]);
let tracked = tracker.update(detections, 1, Some(&transform));

Feature Flags

[dependencies]
norfair = { git = "...", features = ["opencv"] }

Feature	Description
opencv	Enable video I/O, drawing, and homography transforms

License & Attribution

norfair-rs is licensed under the BSD 3-Clause License.

This Rust port is based on the original norfair by Tryolabs (BSD 3-Clause). Their well-designed, detector-agnostic architecture made this port possible. Internal modules include code adapted from several Python libraries—see THIRD_PARTY_LICENSES.md for complete attribution.

Citation: If using this library in research, please cite the original norfair paper as described here.

---

Contributing: Issues and pull requests welcome!