MantisTracker

Rocket Tracker: follow and capture your rocket when launching.

The tracker detects the rocket on a live camera feed and adjust its stand to follow it.

Structure

• /academic: Bachelor report and videos which compare trackers.
• /libraries: Externel libraries used in the project. The references are in the folder.
• /outputs: Images generated by the detector/tracker.
• /src: Implementation of the tracker software.
• /tools: Entry points to run the services.
• /training: Models, config files, trained weights.

Requirements

Install the following packages in order to run the tracker:

For all methods:

• OpenCV (>= 3.4): pip install opencv-contrib-python
• numpy: pip insall numpy
• pyyaml (>= 5.1): pip install pyyaml

If using conda: conda create --name mantis -c anaconda opencv pyyaml numpy

If you have GPU, ensure CUDA/cuDNN are properly installed.

Additionally, if using conda: conda install cudatoolkit=CUDA_VERSION cudnn

For SiamRPN:

Please follow this link to install the requirements: Pysot installation

For ONNX:

• ONNX runtime : pip install onnxruntime

Usage

Before using a detector, please download the weights and configure YOLO in the config.yaml file.

For configuring SiamRPN, the config files are in training/tracking/pysot/. You can add CUDA: False to the config file to disable the use of a GPU.

pipeline.py is the entry point of the tracker. Run python3 tools/pipeline.py with the following parameters:

• --benchmark: use this option with any method to run a benchmark that returns the speed and the accuracy of the run.

• detect: detect a rocket on the provided images

  • --images: path to the images to use in the detection (required).
  • --method: specify the method of detection (default: Yolo). The options are [Yolo, tiny-Yolo].

• track: track a rocket on a provided video feed.

  • --video: the video feed. Can be a path to a video
  • --method: specify the method of tracking (default: OpenTracker). The options are [MOSSE, KCF, SiamRPN, SiamRPN-onnx].
  • --detection: specify the method of detection (default: Yolo). The options are [Yolo, tiny-Yolo].

Examples:

• python3 tools/pipeline.py detect --images training/detection/dataset/test/*.jpg --method tiny-Yolo
• python3 tools/pipeline.py --benchmark track --video training/tracking/dataset/test/test2.mp4 --method SiamRPN

The generated images of the detection or the tracking can be found in the folder /outputs.

Export

tools/export.py is an experimental tool to export the SiamRPN tracker to ONNX and TensorRT.

• python3 tools/export.py onnx: export SiamRPN to ONNX models. Require onnx to be installed.
• python3 tools/export.py tensorrt: export ONNX models to TensorRT engines. Requires TensorRT to be installed.

Installation on Nvidia Jetson Nano

• OpenCV, numpy, pyyaml using pip
• Pytorch: https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-7-0-now-available/72048
• ONNX runtime: https://developer.nvidia.com/blog/announcing-onnx-runtime-for-jetson/

Model Zoo

Weights can be downloaded there: Model Zoo

Two models are available for YOLO and Tiny YOLO which correspond to the two trainings done with different datasets. Second is best.