Person Re-Identification (ReID) Project

Overview

This project focuses on Person Re-Identification (ReID), a critical technology in computer vision and video analysis. ReID aims to recognize and track individuals across multiple non-overlapping camera views, linking or matching images of the same person captured by different cameras within a surveillance system. This project explores various methodologies, datasets, and models to test and improve ReID systems. The primary goal was to implement and evaluate a hand-coded method developed for a school project.

Applications

Security and Surveillance: Track individuals across cameras in public spaces to identify suspects and monitor crowds. Retail and Marketing: Personalize advertising, identify VIP customers, and optimize store layouts based on customer movement. Healthcare and Robotics: Monitor elderly patients, track hospital patients, and assist robot navigation by identifying individuals.

Methodologies

Key Papers

Person Re-Identification Using Color Features and CNN Features

Semantic-Guided Pixel Sampling for Cloth-Changing Person Re-Identification

Occluded Person Re-Identification Via Relational Adaptive Feature Correction Learning (attention Maps)

Person Re-Identification from CCTV Silhouettes Using Generic Fourier Descriptor (Human Silhouettes)

Long-Term Person Re-Identification Based on Appearance and Gait Feature Fusion under Covariate Changes

Night Person Re-Identification and a Benchmark

Models

YOLOv5: Used for accurate and efficient real-time person detection.

ResNet-50: Extracts meaningful features from detected individuals for accurate ReID.

Triplet Loss: Enhances the learning of discriminative features for better ReID performance.

Datasets

iLIDS-VID

The dataset can be found via this link: https://xiatian-zhu.github.io/downloads_qmul_iLIDS-VID_ReID_dataset.html

Features:

320 individuals captured across 2 disjoint camera views

600 image sequences (2 per person)

Variable sequence lengths (23-192 frames, average 73)

Challenges:

Clothing similarities

Lighting and viewpoint variations

Background clutter

MARS

Overview:

1261 unique pedestrians across 6 cameras

Extension of Market-1501 with more challenging aspects

Features:

Frames captured from synchronized cameras for temporal analysis

Emphasis on motion analysis with tracklets

Over 800k images in the training set alone

Data Augmentation

Random changes in brightness, contrast, and saturation of images

Operates in the HSV color space for intuitive color manipulation

Training Parameters

ResNet-50:

Trained for 50 epochs with batches of 32 images

Similarity Network:

Trained using both Adam and SGD optimizers for 500 epochs

Adam: Learning rate of 0.0001

SGD: Learning rate of 0.01 with momentum of 0.9 and decay of 0.0005