Smorodov/Multitarget-tracker

C++

Multiple Object Tracker, Based on Hungarian algorithm + Kalman filter.

multiple-object-trackingkalman-filterhungarian-algorithmcar-trackingcar-countingpeople-trackingabandoned-detectorface-trackingmobilenet-ssdyolokcf

Métricas principais

Crescimento de estrelas

Estrelas

2.4k

Forks

673

Crescimento semanal

—

Issues

1k2k

mar. de 2014mar. de 2016abr. de 2018mai. de 2020jun. de 2022jun. de 2024jul. de 2026

README

Multitarget Tracker

Latest Features

Add YOLOE detection support THU-MIG/yoloe
Add new SOTA: YOLOv26, YOLOv26-obb and YOLOv26-seg models from ultralytics/ultralytics
Add RT-DETRv4 (API similar D-FINE) detection model RT-DETRs/RT-DETRv4
Add D-FINE seg detection model ArgoHA/D-FINE-seg
Add ByteTrack MOT algorithm based on Vertical-Beach/ByteTrack-cpp
Big code cleanup from old style algorithms and detectors: some bgfg detectors, some VOT trackes, Face and Pedestrin detectors, Darknet based backend for old YOLO etc
YOLOv13 detector works with TensorRT! Export pre-trained PyTorch models here (iMoonLab/yolov13) to ONNX format and run Multitarget-tracker with -e=3 example
Instance segmentation model from RF-DETR detector works with TensorRT! Export pre-trained PyTorch models here (roboflow/rf-detr) to ONNX format and run Multitarget-tracker with -e=3 example
New linear assignment algorithm - Jonker-Volgenant / LAPJV algorithm used in scipy as alternative for Hungarian allgorithm
D-FINE detector works with TensorRT! Export pre-trained PyTorch models here (Peterande/D-FINE) to ONNX format and run Multitarget-tracker with -e=3 example
RF-DETR detector works with TensorRT! Export pre-trained PyTorch models here (roboflow/rf-detr) to ONNX format and run Multitarget-tracker with -e=3 example
YOLOv12 detector works with TensorRT! Export pre-trained PyTorch models here (sunsmarterjie/yolov12) to ONNX format and run Multitarget-tracker with -e=3 example

Demo Videos

Detection & Tracking

Documentation

Core Components

1. Object Detectors

Available through CreateDetector function with different detectorType:

Background Subtraction:
- Built-in: VIBE (tracking::Motion_VIBE), SuBSENSE (tracking::Motion_SuBSENSE), LOBSTER (tracking::Motion_LOBSTER)
- OpenCV: MOG2 (tracking::Motion_MOG2)
- OpenCV Contrib: MOG (tracking::Motion_MOG), GMG (tracking::Motion_GMG), CNT (tracking::Motion_CNT)
- Foreground segmentation uses OpenCV contours producing cv::RotatedRect
Deep Learning Models:
- OpenCV DNN module (tracking::DNN_OCV)
- TensorRT-accelerated YOLO (tracking::Yolo_TensorRT)

2. Matching Algorithms

For solving assignment problems:

Hungarian Algorithm (tracking::MatchHungrian) - O(N³) complexity
LAPJV (tracking::MatchBipart) - O(M*N²) complexity
Distance Metrics:
- Center distance (tracking::DistCenters)
- Bounding box distance (tracking::DistRects)
- Jaccard/IoU similarity (tracking::DistJaccard)

3. Trajectory Smoothing

Kalman filters: Linear (tracking::KalmanLinear) and Unscented (tracking::KalmanUnscented)
State models: Constant velocity and constant acceleration
Tracking modes: Position-only (tracking::FilterCenter) and position+size (tracking::FilterRect)
Specialized features: Abandoned object detection, line intersection counting

4. Visual Search

When targets disappear:

KCF (tracking::TrackKCF)
CSRT (tracking::TrackCSRT)
DaSiamRPN (tracking::TrackDaSiamRPN)
Vit (tracking::TrackVit)
Nano (tracking::TrackNano)

Processing Pipelines

Synchronous (SyncProcess): Single-threaded processing
Asynchronous (2 threads) (AsyncProcess): Decouples detection and tracking
Fully Asynchronous (4 threads): For low-FPS deep learning detectors

Installation & Building

git clone https://github.com/Smorodov/Multitarget-tracker.git
cd Multitarget-tracker
mkdir build && cd build
cmake . .. \
  -DUSE_OCV_BGFG=ON \
  -DUSE_OCV_KCF=ON \
  -DUSE_OCV_UKF=ON \
  -DBUILD_ONNX_TENSORRT=ON \
  -DBUILD_ASYNC_DETECTOR=ON \
  -DBUILD_CARS_COUNTING=ON
make -j

Usage Guide

Basic command syntax:

./MultitargetTracker <video_path> [--example=<num>] [--start_frame=<num>] 
                     [--end_frame=<num>] [--end_delay=<ms>] [--out=<filename>]
                     [--show_logs] [--gpu] [--async] [--res=<filename>]
                     [--settings=<filename>] [--batch_size=<num>]

Example:

./MultitargetTracker ../data/atrium.avi -e=1 -o=../data/atrium_motion.avi

Keyboard Controls:

m: Toggle play/pause
Any key: Step forward when paused
Esc: Exit

Integration as Library

#include <mtracking/Ctracker.h>

std::unique_ptr<BaseTracker> m_tracker;
TrackerSettings settings;
settings.SetDistance(tracking::DistJaccard);
m_tracker = BaseTracker::CreateTracker(settings);