Vedelo

Vedelo is a GPU-accelerated vehicle-detection and multi-object-tracking pipeline for video. It fine-tunes and exports YOLO models on the Vedelo Dataset V1, then runs exported CUDA TorchScript models through a Rust runtime. The runtime decodes video, performs batched inference, assigns stable IDs with SORT, draws annotations, and writes an MP4 result.

The repository contains both stages:

• vedelo/ — Python utilities for training, exporting, and model/data downloads.
• runtime/ — the Rust inference and tracking executable, vedelo-rt.
• training-stage/ — expected location for datasets and exported model artifacts.

Requirements

The supported runtime target is x86_64 Debian-based Linux with an NVIDIA GPU and a CUDA-capable PyTorch/libtorch installation. Building requires:

• NVIDIA driver and CUDA support appropriate for the installed PyTorch build.
• build-essential, pkg-config, git, make, and a C compiler toolchain.
• Rust, uv, and just.
• Python 3.14 or newer.

The build compiles pinned FFmpeg, x264, and NVIDIA codec headers from the Git submodules. It uses those libraries for H.264 decode/encode and NVIDIA codec support; a system FFmpeg development package is not used by the just build.

Build

Clone the repository with its submodules, create the Python environment, build the native video dependencies, and compile the release runtime:

git clone --recurse-submodules <repository-url> vedelo
cd vedelo
just venv
just build-deps
just build

just build creates the release binary at runtime/target/release/vedelo-rt and symlinks it to ./vedelo-rt. If the repository was cloned without submodules, run git submodule update --init --recursive before just build-deps.

Run inference

The runtime accepts an exported TorchScript model, an input video, and a destination video. CUDA is selected automatically when available; otherwise it runs on the CPU.

./vedelo-rt \
  --model training-stage/artifacts/exported/v1n_batch8_imgsz1280_best_cuda.torchscript \
  --source training-stage/dataset/test/video_test.mp4 \
  --destination assets/tracked.mp4 \
  --conf 0.6 \
  --iou-thresh 0.1 \
  --max-age 90 \
  --batch-size 8 \
  --profile

Key options:

• --conf is the minimum detection confidence.
• --iou-thresh and --max-age tune SORT track association and lifetime.
• --batch-size is the number of decoded frames evaluated in one model call; increase it only while GPU memory permits. It must be at least 1.
• --imgsz defaults to 1280 and must match the intended model preprocessing.
• --profile prints total elapsed time and end-to-end FPS after encoding.

For the repository’s standard test video and exported models, use the bundled recipes:

just run-v1n
just run-v1s

They write annotated videos under assets/. Ensure the corresponding model artifacts and training-stage/dataset/test/video_test.mp4 are present first.

Measured inference performance

The following end-to-end measurements use the bundled 1,800 × 1,100, 20 FPS test video (202 frames), batch size 8, CUDA TorchScript models, and include decode, preprocessing, inference, tracking, drawing, and output encoding.

Host: NVIDIA GeForce RTX 5060 8 GB and AMD Ryzen 5 5600X.

Model	Confidence	Elapsed	End-to-end throughput	Unique IDs
v1n_batch8_imgsz1280_best_cuda	0.6	10.112s	19.98 FPS	4
v1s_batch8_imgsz1280_best_cuda	0.7	11.307s	17.87 FPS	4

These are workload-specific end-to-end results, not raw model-only benchmark figures. Resolution, codec, storage, GPU driver, model settings, and tracker parameters can change the observed FPS.

Train and export

After setting up the environment, the primary development recipes are:

just download-dataset
just train
just export

just export produces TorchScript and ONNX exports from the configured training artifact. Review the constants in vedelo/train.py and vedelo/export.py before launching a new training run or export.