Deep-SORT

In this work, we implement a popular object tracking algorithm, DeepSORT for tracking thermal images. We have modified the Github repository to track thermal images from the FLIR dataset.

Train

• Thermal object detections were trained in the following repository

Data

In order to be able to test the repository, you will need to download some data first.

• To test the tracker on RGB images, you will need the MOT16 benchmark sequences. This code assumes that the MOT16 benchmark data is in ./MOT16.
• You will also need .npy files which store the detections and their features. The resources folder needs to be extracted to the root directory of this repository.
• The thermal dataset can be found on their website. Make sure the images are stored in ./Thermal/img1/ in the root repository.
• Also download the .npy files for the Thermal dataset and store them in the ./Thermal/ folder.

Run the code

Once you have the above files you can run the tracker with the following command:

python3 deep_sort_app.py \
	--sequence_dir=./Thermal/ \
	--detection_file=./Thermal/seq1.npy \
	--min_confidence=0.8 \
	--nn_budget=100 \
	--display=True