<p>Person Re-ID (Re-ID) is a basic problem in computer vision and smart surveillance, focusing on the association of individuals across distinct camera perspectives. Deep learning has significantly improved person Re-ID performance in recent years, however, the rapid growth of methodologies has made it increasingly challenging to acquire a clear and coherent understanding of the field. In this paper, we provide a thorough and methodical survey of person Re-ID methods based on deep learning, conducted under the PRISMA framework to ensure transparency and reproducibility. We present an innovative multidimensional taxonomy that categorizes existing methodologies along three complementary dimensions: feature representation strategies (global, local, and attribute-based), training paradigms (classification-based and metric learning–based), and network architectures (CNN, Transformer, and GNN-based). We examine the design motivations, advantages, and disadvantages of representative methods in each category within this taxonomy. In addition, we examine well-known image and video-based person Re-ID benchmarks and summarize their data characteristics, specifications, and representative results. Through this systematic analysis, we observe that recent advances are increasingly driven by hybrid learning frameworks, multi-granularity feature modeling, and transformer-based architectures. Additionally, it appears that performance on standard benchmarks is nearing saturation, suggesting the need for new datasets, evaluation settings, and research directions. By synthesizing methodological developments, benchmark evolution, and emerging trends, this survey serves as a comprehensive reference for both new and experienced researchers. It also points out open problems and promising areas for future person Re-ID research.</p>