Overview
This project was developed in collaboration with microbiologists to detect and track bacteria attaching to a glass substrate. In the experimental setup, water containing bioluminescent bacteria flows across the substrate while a microscope records images at 0.5 Hz. Bacteria may attach to the surface and later detach.
The objective was to build a detection and tracking pipeline capable of measuring attachment behavior under different experimental conditions, including the bacteria’s angle relative to the flow.
Procedure
We started with 500 images and applied a basic green-channel threshold to create an initial dataset. We trained a compact YOLOv5 model, then used model-assisted labeling in Labelbox to refine the annotations and improve the dataset.
For tracking, we combined a Kalman filter with bounding-box overlap matching and orientation estimation. The resulting metadata was exported to CSV and analyzed with Pandas.
Skills developed
✅ Object detection
✅ PyTorch
✅ Dataset creation and data analysis
✅ MAL
✅ Communication