A Hybrid Deep Learning-based Person Re-identification Across Multi-Cameras

Surveillance systems play an essential role in enhancing security and ensuring safety within society. Public spaces and critical areas are commonly monitored through multi-camera networks for tracking individuals in real-time to enable quick responses to potential security threats. However, variations in lighting, camera angles, and occlusions pose significant challenges to the process of recognizing individuals across multiple camera views. This research proposes a Person Re-Identification (Re-ID) with deep learning and feature extraction method to improve the accuracy of individual matching and identification despite these challenges. This research encompasses multiple neural network topologies, such as ResNet50, ResNet101, and OSNet architectures with traditional computer vision methodologies such as Scale-Invariant Feature Transform (SIFT) to examine the body part weighting coefficients and feature selection mechanisms. The results demonstrated that integrating dimensionality reduction via principal component analysis (PCA) coupled with semantic segmentation techniques yields statistically significant improvements in re-identification accuracy. The ResNet101 architecture emerges as the superior methodology, achieving performance with a Rank-1 accuracy of 0.88 and mean Average Precision (mAP) of 0.94 on the test dataset.