A C++17 tracking switchboard wired for four famous algorithms

What it does

DeepTracker is a real-time multi-object tracking application built in C++17 with Qt 6 and ONNX Runtime. It runs a YOLO detector through one of four pluggable trackers—ByteTrack, OC-SORT, Deep OC-SORT, or DeepSORT—behind a thin ITracker interface. A Qt desktop app provides interactive tracking with live preview and trajectory trails, while a command-line demo handles headless use.

The interesting bit

The project treats tracker choice as a policy decision rather than an architectural one. ByteTrack, OC-SORT and Deep OC-SORT all inherit from a shared BaseTracker template-method pipeline and override virtual hooks to change association behavior, while DeepSORT stands alone with its own faithful age-ordered cascade, 100-frame feature gallery, and χ² motion gating. Switching between them is nominally a one-line change.

Key highlights

• All four trackers share the same ITracker interface (update, reset, frame_id) and a constant-velocity Kalman filter on the [cx, cy, aspect, height] state.
• Detection and ReID embeddings run through ONNX Runtime with plain .onnx files, so models can be swapped without recompiling; inference defaults to CPU but can target CUDA, DirectML, or OpenVINO without touching tracking code.
• DeepSORT falls back to ByteTrack automatically when no ReID model is present, and Deep OC-SORT likewise degrades to plain OC-SORT without appearance features.
• The Qt GUI discovers models automatically, supports color-theme switching, and exposes Global Motion Compensation options (sparseOptFlow, orb, sift, ecc) for moving-camera footage.
• Association math includes a Jonker–Volgenant linear assignment solver and both IoU and embedding distance matrices.

Caveats

• Tracker parameters are compiled-in presets inside src/core/tracker_config.cpp; tuning behavior requires editing C++ and rebuilding.
• The build system is Windows-centric: it expects Visual Studio 2022, Qt 6.9.x, and prebuilt OpenCV and ONNX Runtime libraries placed under external/.
• The README does not mention Linux or macOS build instructions, and the .pro file paths suggest a Windows-first deployment.

Verdict

Worth a look if you need a self-contained C++ tracker with algorithmic variety and a Qt frontend. Skip it if you were hoping for runtime parameter tuning or a cross-platform build out of the box.