Feature Extraction Algorithm for Person Re-Identification Based on Artificial Intelligence

Re-IDentification (Re-ID) of persons is significant in intelligent surveillance and security systems, where the consistency of feature extraction and prediction processes is essential to ensuring correct identity matches. Illumination, camera view, and visual quality differences often cause noise in pseudo-labels in unsupervised Person Re-ID tasks, which undermines the overall learning performance and enhances the quality of the image before feature extraction, creating a gap in the process that impacts the reliability of representations and prediction accuracy. An artificial intelligence (AI)-driven Person Re-ID framework is proposed to overcome this challenge. Pedestrian multi-camera Re-ID dataset containing approximately 68,000 pedestrian images of 1,501 identities captured from multiple surveillance cameras. Multi-scale Retinex-based illumination correction and edge-aware detail sharpening preprocessing techniques enhance the quality of images and provide a stable input image. The Scale-Invariant Feature Transform (SIFT) is used to extract features and offers cross-view pedestrian matching features, as it is an invariant and stable feature extractor. The Namib Beetle Optimized Residual Neural Network (NBO-ResNet) acts as the prediction model and enhances the classification strength by applying biologically inspired optimization strategies. Findings show a definite increase in Rank-1 accuracy (97.3%), a mAP (95.8%), the stability of clustering, and the consistency of pseudo-labels, which proves the ability of the framework to operate effectively in unsupervised settings. Python is used to implement the framework on the PyTorch deep learning (DL) platform. The proposed framework enhances unsupervised person Re-ID by improving data quality, feature representation, and prediction reliability.