Projects

Person Re-Identification Dissertation

1 April 2025

Python
YOLOv11
ResNet-50
OpenCV
PyTorch
Computer Vision

A hybrid person re-identification system combining YOLO-based segmentation with colour analysis to improve accuracy under complex visual conditions. Evaluated against ResNet-50 using a custom ethical dataset, the system demonstrated greater robustness and precision across varied image environments.

Clothing item segmenation visualisation
Detection of clothing items with confidence
Comparison of clothing items of same person in two different images to get overall dissimilarity score
Comparison of clothing items of different people, displaying the dissimilarity score going up
Top 5 matches visualised from a query image against the entire gallery. Including an incorrect result
Projects test coverage displayed
Class diagram of project

Description:
My dissertation introduced a novel hybrid approach to person re-identification (Re-ID) that integrates YOLO object segmentation with colour histogram analysis to enhance identification accuracy in challenging, real-world scenarios. The system was implemented in Python, benchmarked against ResNet-50, and tested using an ethically generated synthetic dataset to ensure privacy compliance. Results showed the hybrid method offered improved mean Average Precision (mAP) and greater robustness to background noise and lighting variability, highlighting its potential for scalable, privacy-conscious Re-ID applications such as smart city systems or missing person detection.

Links: