Multi Object Tracking

Yet another C++ adaptation of SOTA multi-object tracking algorithm

Supported Trackers

• SORT
• BoTSORT

Dataset

Get the MOT dataset of your choice from MOT website

cd data
wget https://motchallenge.net/data/MOT20.zip
unzip MOT20.zip && rm MOT20.zip

Configure

In config folder, add your custom tracker config

SORT

tracker = "sort"
max_time_lost = 15
match_thresh = 0.3

[kalman]
time_step = 1
process_noise_scale = 1.0
measurement_noise_scale = 1.0

BoTSORT

tracker = "botsort"
max_time_lost = 15
track_high_thresh = 0.5
track_low_thresh = 0.1
new_track_thresh = 0.6
first_match_thresh = 0.3
second_match_thresh = 0.1
unconfirmed_match_thresh = 0.2
proximity_thresh = 0.5
appearance_thresh = 0.9

[kalman]
time_step = 1
process_noise_scale = 1.0
measurement_noise_scale = 1.0

🖥️ Local

Compile

# Ensure to reset the deps
meson subprojects update --reset
meson setup build --wipe

meson compile -C build
meson test -C build

Run

cd build/app
./mot -h

# Example
./mot -i data/MOT20/train/<seq-name> -c config/sort.toml --display

Evaluate

First, set up your Python environment and dependencies:

python3 -m venv venv
./venv/bin/pip install -r requirements.txt

Then run the evaluation script:

chmod +x mot-eval.sh
./mot-eval.sh --dataset data/MOT20 --split train --config app/config/sort.toml --save
# experiment output available in runs folder

---

🐳 Docker

Build the image:

docker build -t mot.cpp .

Run the tracker on a dataset sequence mounted from the host:

# Allow Docker to connect to the X11 display server
xhost +local:docker

docker run --rm \
    --user $(id -u):$(id -g) \
    --env DISPLAY=$DISPLAY \
    -v /tmp/.X11-unix:/tmp/.X11-unix:ro \
    -v $(pwd)/data:/data \
    -v $(pwd)/app/config:/opt/mot.cpp/app/config \
    mot.cpp -i /data/MOT20/train/MOT20-01 -c /opt/mot.cpp/app/config/sort.toml --display

Run evaluation across a full dataset split:

docker run --rm \
    --user $(id -u):$(id -g) \
    -v $(pwd)/data:/data \
    -v $(pwd)/runs:/opt/mot.cpp/runs \
    --entrypoint /opt/mot.cpp/mot-eval.sh \
    mot.cpp --dataset /data/MOT20 --split train --config /opt/mot.cpp/app/config/sort.toml --save

Run with your detector

https://github.com/tensorworksio/TensorRT-Vision/tree/main/app/mot