A multi-object tracking framework featuring YOLOv8 detection, DeepSORT-based association, and three evolutionary optimizers (QPSO, NSGA-II, MOPSO). Designed for robust benchmarking, automatic tuning, and performance visualization across multiple feature extractors and tracking configurations.

• Track-by-detection framework using YOLOv8 + DeepSORT

• GPU-accelerated inference via PyTorch & Ultralytics

• Three evolutionary optimizers:

  • QPSO (Quantized Particle Swarm Optimization)
  • NSGA-II (Non-dominated Sorting Genetic Algorithm)
  • MOPSO (Multi-objective PSO)

• Three feature extractors (embedders):

  • MobileNetV2
  • ShuffleNetV2
  • ResNet

• Automatic batch optimization across all 3×3 (embedder × optimizer) combinations

• Mock MOTA, IDF1, and FPS evaluation for fast prototyping

• Pareto front visualizations (2D and 3D)

• Modular codebase for easy extension & full GPU compatibility

NPP-People-Tracking/ ├── detectors/ # YOLOv8 detection wrapper ├── embeddings/ # Feature extractors ├── evaluation/ # Evaluation logic ├── optimization/ # QPSO, NSGA, MOPSO + plotting ├── results/ # Logs, metrics, plots ├── sample_videos/ # Input test videos ├── trackers/ # DeepSORT / ByteTrack wrappers ├── utils/ # Drawing and utility helpers ├── main.py # Visual demo runner ├── batch_runner.py # Full automation across all configs ├── run_plot_pareto.py # Manual plotting script ├── config.yaml # Tuning bounds for all optimizers 

pip install -r requirements.txt

wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8n.pt

Place it in the root directory as yolov8n.pt

Place your input video in sample_videos/test_video.mp4

python batch_runner.py

This will:

• Run QPSO, NSGA-II, and MOPSO for MobileNetV2, ShuffleNetV2, and ResNet
• Save logs, metrics, plots
• Output a combined summary CSV at results/summary_table.csv

python main.py

Visualize tracking with live FPS overlay and bounding boxes.

python run_plot_pareto.py

• metrics.json: Best config + result for each combo
• log.csv: All 50 generations of evaluations
• *_2d.png, *_3d.png: Saved Pareto plots (FPS vs MOTA, FPS vs MOTA vs IDF1)
• summary_table.csv: Global table across all runs

• All tracking logic and evaluation runs on GPU by default
• Mock metrics are used for rapid testing, customizable in evaluation.py
• Optimizers operate over bounds defined in optimization/config.yaml
• Supports future upgrades for real GT-based evaluation using py-motmetrics

• Abhinav Shukla Project design, optimization logic, performance benchmarking