YOLO Object Detection & Tracking
Real-time object detection and multi-object tracking experiments comparing YOLO with several tracking algorithms.
- Computer Vision
- Robotics
TL;DR
Experiments combining YOLO detection with multiple tracking methods (ByteTrack, OC-SORT, DeepSORT), benchmarking accuracy/speed trade-offs with deployment in mind. (Placeholder narrative.)
Problem
Detection alone isn't enough for video — you need stable identities across frames. The goal was to find a detector+tracker combo that is accurate and real-time.
Note: Placeholder case-study content — replace with your real write-up.
Why it matters
Most real-world vision tasks happen in video, where you need consistent object identities over time. Choosing the right tracker is a balance of accuracy, speed, and robustness to occlusion.
Dataset / inputs
Video clips with detection/tracking ground truth. (Placeholder — name the clips/benchmark and hardware used.)
Technical decisions
I held the detector fixed (YOLO) and swapped trackers to isolate their effect. ByteTrack and OC-SORT are fast and motion-based; DeepSORT adds appearance features that cut ID switches at a higher compute cost.
Challenges
Occlusions and crowded scenes caused most ID switches. Keeping the benchmark fair — same clips, same metrics — was essential to drawing honest conclusions.
Methods
- YOLO detector as the front end
- ByteTrack / OC-SORT / DeepSORT compared as trackers
- Benchmarking on shared clips with consistent metrics
Results
- Speed/accuracy comparison across trackers (placeholder)
- Notes on when each tracker is preferable (placeholder)
Lessons learned
- Tracker choice depends heavily on FPS budget and occlusion frequency
- Appearance-based trackers (DeepSORT) cost more but reduce ID switches
Limitations
- Evaluated on a limited set of clips
- No edge-device deployment yet
Next steps
- Deploy to an edge device and re-benchmark
- Test under heavier occlusion and crowd density
Related projects
Gap Junction Segmentation in Electron Microscopy
Deep-learning segmentation pipeline for detecting thin gap-junction structures in electron-microscopy volumes.
Outcome: Example result to update — improved recall on thin structures vs. a baseline U-Net.
- Biomedical AI
- Computer Vision
- Research
Face Recognition Attendance System
Team-built attendance system using face recognition, with stranger rejection, reporting, and a dashboard.
Outcome: Example result to update — automated attendance capture with controlled false-positive rate.
- Computer Vision
- Applied AI
- MLOps / Deployment