shaoshengsong/DeepSORT

A real-time multi-object tracking application in C++17 with Qt and ONNX Runtime that runs a YOLO detector with four pluggable trackers — ByteTrack, OC-SORT, Deep OC-SORT, and DeepSORT.

What it solves

DeepTracker is a real-time multi-object tracking (MOT) system for video. It allows a user to identify and maintain the identity of multiple objects (like people) across video frames, handling challenges such as occlusions, motion blur, and moving cameras.

How it works

The system uses a YOLO detector to find objects in each frame. It then applies one of four tracking algorithms to associate these detections with existing tracks. All trackers share a common C++ interface (ITracker) and a constant-velocity Kalman filter for motion prediction.

Depending on the chosen algorithm, the system associates objects using different methods:

  • ByteTrack: Uses motion (IoU and confidence) to recover tracks lost to occlusion.
  • OC-SORT: Focuses on motion (observation-centric) to handle blur and occlusion without needing appearance models.
  • Deep OC-SORT: Combines motion and appearance features, adding Global Motion Compensation (GMC) for moving cameras.
  • DeepSORT: Uses a combination of appearance embeddings (ReID) and squared-Mahalanobis motion gating to maintain identities.

Inference is handled via ONNX Runtime, allowing the use of .onnx models for detection and appearance encoding without needing to be recompiled.

Who it’s for

This project is for developers and researchers who need a high-performance C++ implementation of classic multi-object tracking algorithms for video analysis.

Highlights

  • Multiple Algorithms: Supports ByteTrack, OC-SORT, Deep OC-SORT, and DeepSORT.
  • Flexible Inference: Uses ONNX Runtime to support CPU, CUDA, DirectML, and OpenVINO providers.
  • Harnesses YOLO: Integrates a YOLO detector for the initial object detection phase.
  • Interactive GUI: Includes a Qt-based desktop application for interactive tracking and a command-line tool for scripting.
  • Modular Architecture: All trackers implement a shared interface, making it easy to switch between algorithms with a single line of change.

Related

  • Project
    jayin92/Skyfall-GS

    Skyfall-GS is a framework that synthesizes large-scale, immersive 3D urban scenes by combining satellite imagery for geometry and diffusion models for high-quality textures.

  • Project
    Jumpat/SegAnyGAussians

    SAGA is a framework for segmenting 3D Gaussian Splatting scenes, allowing users to isolate 3D objects using interactive point prompts or open-vocabulary text queries.

  • Project
    nghorbani/amass

    A unified database and toolkit for human motion capture, consolidating various marker-based datasets into a common framework for animation and deep learning training.

  • Project
    microsoft/aurora

    Aurora is a foundation model for the Earth system that predicts atmospheric variables like temperature, air pollution, and ocean waves by adapting a general-purpose pre-trained model to specialized forecasting tasks.