Project: Advanced Multi-Object Tracking with StrongSORT + YOLOv8

State-of-the-art multi-object tracking pipeline using YOLOv8m detection and StrongSORT with OSNet appearance embeddings, evaluated on MOT17 against a ByteTrack baseline.

Built for video analytics and surveillance roles — demonstrates track lifecycle management, appearance-based re-association, camera motion compensation, and formal evaluation with HOTA/MOTA/IDF1 metrics.

---

Results

StrongSORT (Primary)

Sequence	HOTA	MOTA	MOTP	IDF1	IDSw
MOT17-02-SDP	31.18	24.30	81.75	34.83	69
MOT17-04-SDP	43.06	37.49	82.86	52.59	58
MOT17-05-SDP	48.76	51.02	76.97	65.91	66
MOT17-09-SDP	51.39	64.26	81.98	62.61	41
MOT17-10-SDP	39.60	42.88	76.34	49.48	103
COMBINED	41.60	38.15	80.89	50.79	337

StrongSORT vs ByteTrack

Sequence	HOTA SS/BT/Δ	MOTA SS/BT/Δ	IDF1 SS/BT/Δ	IDSw SS/BT
MOT17-02-SDP	31.2/30.6/+0.6	24.3/22.9/+1.4	34.8/33.5/+1.3	69/25
MOT17-04-SDP	43.1/39.2/+3.9	37.5/31.9/+5.6	52.6/45.4/+7.2	58/27
MOT17-05-SDP	48.8/43.8/+4.9	51.0/50.8/+0.2	65.9/59.6/+6.3	66/88
MOT17-09-SDP	51.4/49.6/+1.8	64.3/64.4/-0.1	62.6/62.0/+0.6	41/27
MOT17-10-SDP	39.6/34.5/+5.1	42.9/39.7/+3.2	49.5/44.0/+5.5	103/76
COMBINED	41.6/38.2/+3.4	38.1/34.5/+3.7	50.8/45.5/+5.3	337/243

StrongSORT outperforms ByteTrack on HOTA, MOTA, and IDF1 across all sequences. ByteTrack produces fewer ID switches on most sequences — a consequence of StrongSORT's aggressive appearance-based re-association, which recovers more lost tracks but occasionally re-assigns incorrectly.

---

Tech Stack

• Detector: YOLOv8m (pretrained, no fine-tuning)
• Tracker: StrongSORT with OSNet-x0.25 appearance embeddings (boxmot)
• Baseline: ByteTrack (boxmot)
• Evaluation: TrackEval (HOTA, MOTA, MOTP, IDF1)
• Dataset: MOT17 — 5 sequences (train split)
• Environment: conda, Python 3.10, PyTorch 2.x, M1 MacBook Air

---

Project Structure

├── config/
│   └── config.yaml          # All parameters — no hardcoded values
├── src/
│   ├── data_loader.py       # MOT17 frame iterator with resolution scaling
│   ├── detector.py          # YOLOv8m wrapper
│   ├── tracker.py           # StrongSORT + lifecycle state management
│   ├── tracker_byte.py      # ByteTrack wrapper (same interface)
│   ├── visualizer.py        # Bounding boxes, state labels, velocity vectors
│   ├── video_writer.py      # MP4 export
│   ├── reporter.py          # Per-sequence JSON analytics
│   └── eval_formatter.py   # MOT Challenge format exporter
├── scripts/
│   ├── run_tracking.py      # Full StrongSORT pipeline across all sequences
│   ├── run_baseline.py      # ByteTrack pipeline
│   ├── run_eval.py          # TrackEval runner (--tracker flag)
│   └── compare_trackers.py  # Side-by-side comparison table
├── outputs/
│   ├── videos/              # Annotated MP4s (gitignored)
│   ├── reports/             # Per-sequence JSON analytics (gitignored)
│   └── eval/                # TrackEval results and comparison
└── data/                    # MOT17 dataset (gitignored)

---

Setup

git clone https://github.com/tajwarchy/advanced-multi-object-tracking
cd advanced-multi-object-tracking

conda create -n mot_tracking python=3.10 -y
conda activate mot_tracking

pip install torch torchvision torchaudio
pip install ultralytics boxmot opencv-python pyyaml

# TrackEval
git clone https://github.com/JonathonLuiten/TrackEval.git ~/TrackEval
cd ~/TrackEval && pip install -e . && cd -

Update eval.trackeval_root in config/config.yaml if TrackEval is cloned elsewhere.

---

Dataset Preparation

Download MOT17 from https://motchallenge.net/data/MOT17.zip and extract:

unzip MOT17.zip -d data/

Set up GT symlinks for TrackEval:

mkdir -p outputs/eval/gt/MOT17/MOT17-train
for seq in MOT17-02-SDP MOT17-04-SDP MOT17-05-SDP MOT17-09-SDP MOT17-10-SDP; do
    ln -s "$(pwd)/data/MOT17/train/$seq" \
          "$(pwd)/outputs/eval/gt/MOT17/MOT17-train/$seq"
done

---

Running

# StrongSORT — full pipeline (all 5 sequences)
python scripts/run_tracking.py

# ByteTrack baseline
python scripts/run_baseline.py

# Evaluate StrongSORT
python -m scripts.run_eval --tracker StrongSORT

# Evaluate ByteTrack
python -m scripts.run_eval --tracker ByteTrack

# Side-by-side comparison
python scripts/compare_trackers.py

---

Visual Outputs

Each annotated video includes:

• Color-coded bounding boxes per track ID (consistent across frames)
• State label overlay: [B] born, [A] active, [L] lost
• Velocity vectors on moving targets (suppressed below noise threshold)

Per-sequence JSON reports in outputs/reports/ include track IDs, lifetimes, trajectory centroids, and inference FPS.

---

Environment Notes

• M1 MacBook Air — MPS for YOLOv8m inference, CPU fallback for tracker if MPS is unstable
• num_workers=0 throughout (macOS multiprocessing constraint)
• Inference resolution: 640px (reduce to 480px if thermal throttling occurs)
• Coordinates rescaled to original resolution before TrackEval submission