zghias/Multiple-Object-Tracking-with-SORT

Multiple Object Tracking with SORT and DeepSORT

Author: Zuhair Ghias

Email: zuhairg2@illinois.edu

Source code: https://github.com/ZuhairGhias/Multiple-Object-Tracking-with-SORT

Demo: https://huggingface.co/spaces/zghias/Multiple-Object-Tracking-with-SORT

This project is an interactive technical demo for multiple object tracking on MOT17. It compares a progression of tracking-by-detection methods:

• —Naive IoU association
• —SORT
• —DeepSORT
• —MyDeepSORT2, a lightweight DeepSORT variant using HSV color histograms and IoU-gated association

The Gradio app includes curated MOT17-09-SDP videos, metric plots, runtime plots, and a written report explaining the methods and tradeoffs.

Demo

Run the app locally:

pip install -r requirements.txt
python app.py

The app opens at a local Gradio URL. The demo videos are stored under:

data/videos/MOT17-09-SDP/

Included videos:

• —source.mp4
• —detections.mp4
• —tracking_naive_iou.mp4
• —tracking_sort.mp4
• —tracking_deep_sort.mp4
• —tracking_my_deep_sort2.mp4
• —tracking_gt.mp4

Project Structure

Core source code:

• —src/methods/tracking/SORT.py: SORT implementation
• —src/methods/tracking/deep_SORT.py: DeepSORT implementation and shared DeepSORT helpers
• —src/methods/tracking/MyDeepSORT2.py: lightweight color-histogram DeepSORT variant
• —src/metrics/: MOT metric calculation helpers
• —src/utils/mot17.py: MOT17 video generation utility
• —src/utils/mot17_metrics.py: MOT17 metrics pipeline
• —src/utils/mot17_metrics_plots.py: static plot and table generation
• —app.py: Gradio demo and written report

Generating Demo Videos

The committed demo uses MOT17-09-SDP. To regenerate those videos from a local MOT17 checkout:

python -m src.utils.mot17 MOT17-09-SDP --output-dir data/videos

The video utility writes source, detections, ground truth, Naive IoU, SORT, DeepSORT, and MyDeepSORT2 MP4s. It uses H.264 output when imageio-ffmpeg is installed, which improves browser playback in Gradio.

Generating MOT17 Metrics

The metrics utility scores every locally available MOT17 training sequence under data/MOT17/train that contains:

• —seqinfo.ini
• —det/det.txt
• —gt/gt.txt

Run:

python -m src.utils.mot17_metrics

The command writes:

data/metrics/MOT17_tracking_metrics.csv

The CSV contains independent sequence rows, detector/tracker aggregate rows, overall aggregate rows, frame counts, prediction counts, runtime fields, and standard MOT metrics such as MOTA, MOTP, IDF1, FP, FN, ID switches, and fragmentations.

Plotting Metrics

Generate the report plots from the saved CSV:

python -m src.utils.mot17_metrics_plots

The app currently uses the full comparison plots for:

• —Naive IoU
• —SORT
• —DeepSORT
• —MyDeepSORT2

The plot filenames follow this convention:

MOT17_tracking_trackers-{trackers}_detectors-{detectors}_{plot_name}.png

You can generate smaller comparison plots with filters:

python -m src.utils.mot17_metrics_plots --trackers naive_iou,sort
python -m src.utils.mot17_metrics_plots --trackers naive_iou,sort,deep_sort --detectors SDP

Relevant Papers

• —Bewley et al., "Simple Online and Realtime Tracking", 2016. https://arxiv.org/abs/1602.00763
• —Wojke et al., "Simple Online and Realtime Tracking with a Deep Association Metric", 2017. https://arxiv.org/abs/1703.07402
• —Du et al., "StrongSORT: Make DeepSORT Great Again", 2022. https://arxiv.org/abs/2202.13514

AI Use Policy

I worked on the core tracker implementations and evaluation logic. I used AI assistance for brainstorming, utility code, automation, the evaluation framework, debugging, cleanup, and generating repeatable plots and videos. I reviewed and integrated the changes myself.